CN117095005A - Plastic master batch quality inspection method and system based on machine vision - Google Patents

Abstract

The invention relates to the technical field of image analysis, in particular to a plastic master batch quality inspection method and system based on machine vision, comprising the following steps: based on a plastic master batch sample, a multi-sensor synchronous acquisition technology is adopted, and simultaneously visible light, infrared and ultraviolet spectrogram images are acquired to generate a multi-mode original image data set. In the invention, the multi-sensor synchronous acquisition technology integrates visible light, infrared and ultraviolet images, expands detection dimension, improves image quality and detail by a convolution neural network and multi-source image fusion algorithm, creates a high-definition comprehensive image, processes point cloud data to construct a three-dimensional model, visually displays the form of plastic master batches, improves defect detection accuracy by a deep convolution neural network, reduces false alarm, optimizes production parameters in real time by deep reinforcement learning, enhances self-adaptive anomaly detection, supports dynamic image acquisition parameter optimization by reinforcement learning, and improves image quality and detection accuracy.

Description

Plastic master batch quality inspection method and system based on machine vision

Technical Field

The invention relates to the technical field of image analysis, in particular to a plastic master batch quality inspection method and system based on machine vision.

Background

Machine vision is a technique that allows computers and other vision equipment to "see" and interpret their surroundings. In the field of image analysis, this typically involves using cameras and computer algorithms to analyze images and extract useful information therefrom. Image analysis can be used in a variety of fields such as industrial inspection, medical image analysis, environmental perception of autopilot vehicles, etc.

The quality inspection method of the plastic master batch based on machine vision is a method for inspecting the quality of the plastic master batch by utilizing a machine vision technology. The method aims to improve the monitoring and control of the quality of the plastic master batch in the production process and ensure that the produced plastic master batch meets the preset quality standard. The effect that reaches is through machine vision system, carries out high-speed, high accuracy to characteristics such as outward appearance, colour, shape, size, surface defect of plastics master batch. The method can realize automatic detection of a large number of plastic master batches, improves production efficiency, reduces labor cost and reduces quality control errors.

Existing plastic master batch quality inspection methods often rely on only a single sensor for image acquisition, resulting in possible omission of certain defects in a particular band. The method does not have advanced image fusion technology, so that the details and the quality of the image are limited, and high-precision detection is difficult. The lack of modeling and analysis of three-dimensional morphology has a dead zone for defects of complex morphology. Conventional defect detection algorithms often do not fully utilize deep learning techniques, resulting in relatively low recognition and accuracy rates. In addition, the existing method lacks intelligent adjustment capability for production parameters, and when abnormality is detected, manual intervention is needed, so that production efficiency is reduced. Finally, the fixed setting of the image acquisition parameters results in poor image quality in some cases, and it is difficult to meet the requirement of high-precision detection.

Disclosure of Invention

The invention aims to solve the defects in the prior art, and provides a plastic master batch quality inspection method and system based on machine vision.

In order to achieve the above purpose, the present invention adopts the following technical scheme: a plastic master batch quality inspection method based on machine vision comprises the following steps:

s1: based on a plastic master batch sample, adopting a multi-sensor synchronous acquisition technology, and simultaneously acquiring visible light, infrared and ultraviolet spectrogram images to generate a multi-mode original image data set;

s2: based on the multi-mode original image data set, performing image preprocessing by adopting a convolutional neural network, and generating a comprehensive fusion image through a multi-source image fusion algorithm;

s3: based on the comprehensive fusion image, a three-dimensional model of the plastic master batch is established by adopting a point cloud data processing method, and shape and surface analysis is carried out to generate a three-dimensional morphological analysis report;

s4: based on the comprehensive fusion image and the three-dimensional morphological analysis report, performing defect detection and classification on the plastic master batch by adopting a deep convolutional neural network to generate an automatic defect labeling chart;

s5: based on the automatic defect labeling diagram, adopting a deep reinforcement learning algorithm to dynamically adjust production parameters, and carrying out self-adaptive anomaly detection and strategy updating to generate a quality anomaly report and an optimized detection strategy;

S6: and dynamically optimizing image acquisition parameters, including exposure time, focal length and light source intensity, by adopting a reinforcement learning method based on the quality anomaly report, and generating optimized image acquisition parameter settings.

As a further scheme of the invention, based on a plastic master batch sample, a multi-sensor synchronous acquisition technology is adopted, and simultaneously visible light, infrared and ultraviolet spectrogram images are acquired, and the steps for generating a multi-mode original image data set are specifically as follows:

s101: based on the selective strategy, adopting a sensor selection algorithm to perform optimal sensor configuration and generating sensor parameter configuration;

s102: based on the sensor parameter configuration, a synchronous trigger mechanism is applied to synchronously start a plurality of groups of sensors to obtain synchronous acquisition control signals;

s103: based on the synchronous acquisition control signal, adopting a high-speed imaging technology to acquire a visible light image, and generating a visible light original image;

s104: based on the synchronous acquisition control signal, an infrared imaging technology is utilized to acquire an infrared image, and an infrared original image is generated;

s105: based on the synchronous acquisition control signal, an ultraviolet imaging technology is used for acquiring an ultraviolet image, and an ultraviolet original image is generated;

S106: based on the visible light original image, the infrared original image and the ultraviolet original image, a data integration technology is applied, and the multi-mode images are combined to generate a multi-mode original image data set.

As a further scheme of the invention, based on the multi-mode original image data set, a convolutional neural network is adopted to perform image preprocessing, and a multi-source image fusion algorithm is adopted to generate a comprehensive fusion image, wherein the steps of the method specifically comprise:

s201: based on the multi-mode original image data set, adopting an image enhancement algorithm to improve the image quality and obtaining an enhanced image data set;

s202: based on the enhanced image dataset, applying a median filtering algorithm to perform image noise reduction and generating a noise-reduced image dataset;

s203: based on the noise-reduced image data set, performing spatial alignment by adopting an image registration algorithm to obtain a spatially aligned image data set;

s204: performing brightness and contrast adjustment based on the spatially aligned image dataset by using a histogram equalization technique to generate a contrast-adjusted image dataset;

s205: based on the image dataset with the adjusted contrast, extracting key features by using a SIFT feature extraction algorithm to obtain a key image feature set;

S206: and based on the key image feature set, applying a multi-source image fusion algorithm, and combining the multi-mode images to generate a comprehensive fusion image.

As a further scheme of the invention, based on the comprehensive fusion image, a point cloud data processing method is adopted to establish a three-dimensional model of the plastic master batch, shape and surface analysis is carried out, and the step of generating a three-dimensional morphological analysis report specifically comprises the following steps:

s301: based on the comprehensive fusion image, adopting a three-dimensional reconstruction algorithm to perform three-dimensional reconstruction to obtain a preliminary three-dimensional model;

s302: based on the preliminary three-dimensional model, performing model gridding by utilizing a three-dimensional grid generation technology to generate a gridding three-dimensional model;

s303: based on the meshed three-dimensional model, applying a point cloud filtering algorithm to perform data noise reduction to obtain a noise-reduced three-dimensional model;

s304: based on the three-dimensional model after noise reduction, adopting a surface fitting technology to analyze the shape and the surface, and generating three-dimensional shape and surface data;

s305: based on the three-dimensional shape and the surface data, extracting and analyzing a specific form by using a morphological operation technology to obtain a form characteristic report;

s306: integrating the morphological feature report with the three-dimensional shape and surface data, and generating a three-dimensional morphological analysis report by using a data integration technology.

As a further scheme of the invention, based on the comprehensive fusion image and the three-dimensional morphological analysis report, the depth convolution neural network is adopted to detect and classify the defects of the plastic master batch, and the steps for generating the automatic defect labeling chart are specifically as follows:

s401: based on the comprehensive fusion image, adopting an image enhancement algorithm to perform image preprocessing to generate an enhanced plastic master batch image;

s402: based on the reinforced plastic master batch image, performing feature extraction by adopting a SIFT technology to obtain a plastic master batch feature set;

s403: based on the plastic master batch characteristic set, performing defect identification by adopting a deep convolutional neural network to obtain a defect identification tag;

s404: and combining the defect identification label and the reinforced plastic master batch image, and adopting a classification algorithm to classify defects to generate an automatic defect labeling chart.

As a further scheme of the invention, based on the automatic defect labeling diagram, a deep reinforcement learning algorithm is adopted to dynamically adjust production parameters, and self-adaptive anomaly detection and strategy updating are carried out, and the steps of generating a quality anomaly report and an optimized detection strategy are specifically as follows:

s501: analyzing defect distribution by adopting a data analysis technology based on the automatic defect labeling diagram to obtain a defect analysis report;

S502: based on the defect analysis report, adopting a deep reinforcement learning algorithm to adjust production parameters and generating adjusted production parameters;

s503: monitoring a production process by using the adjusted production parameters and adopting a self-adaptive abnormality detection technology to generate an abnormality detection report;

s504: and optimizing the detection strategy by adopting a strategy updating algorithm according to the abnormal detection report to obtain a quality abnormal report.

As a further scheme of the invention, based on the quality anomaly report, the image acquisition parameters are dynamically optimized by adopting a reinforcement learning method, including exposure time, focal length and light source intensity, and the steps for generating the optimized image acquisition parameter settings are specifically as follows:

s601: according to the quality abnormality report, adopting an analysis technology to discuss the abnormality cause and obtain an abnormality cause analysis report;

s602: based on the abnormal cause analysis report, applying a reinforcement learning method, and providing an image acquisition optimization scheme to generate a parameter optimization scheme;

s603: combining the parameter optimization scheme, adjusting exposure time, focal length and light source intensity, and generating primarily optimized image acquisition parameters;

s604: carrying out image acquisition again by utilizing the primarily optimized image acquisition parameters, evaluating the effect of the image acquisition parameters and generating an effect evaluation report;

S605: and according to the effect evaluation report, carrying out parameter refinement and perfecting to generate optimized image acquisition parameter setting.

The plastic master batch quality inspection system based on machine vision is used for executing the plastic master batch quality inspection method based on machine vision, and comprises an image acquisition module, an image processing module, a three-dimensional modeling module, a characteristic recognition module, a quality analysis module, a parameter adjustment module and an optimization acquisition module.

As a further scheme of the invention, the image acquisition module acquires visible light, infrared and ultraviolet images by adopting a synchronous trigger mechanism based on a sensor selection algorithm to generate a multi-mode original image data set;

the image processing module generates an image dataset with improved quality by adopting an image enhancement algorithm, a median filtering algorithm, an image registration algorithm and a histogram equalization technology based on the multi-mode original image dataset, and finally generates a comprehensive fusion image;

the three-dimensional modeling module generates a meshed three-dimensional model and three-dimensional shape surface data by utilizing a three-dimensional reconstruction algorithm and a three-dimensional mesh generation technology based on the comprehensive fusion image;

The feature recognition module is used for preprocessing the comprehensive fusion image, extracting features, performing defect recognition by using a deep convolutional neural network, and generating a defect recognition tag;

the quality analysis module classifies defects based on the defect identification labels, generates an automatic defect labeling chart, analyzes defect distribution and generates a defect analysis report;

the parameter adjustment module adjusts production parameters based on the defect analysis report by using a deep reinforcement learning algorithm, generates adjusted production parameters, monitors the production process, generates an abnormal detection report, optimizes the detection strategy by a strategy updating algorithm, and finally generates a quality abnormal report;

the optimization acquisition module analyzes the abnormal reasons, proposes an optimization scheme, generates a parameter optimization scheme, adjusts image acquisition parameters, generates primarily optimized image acquisition parameter settings, re-acquires images, evaluates effects and finally obtains the optimized image acquisition parameter settings.

As a further scheme of the invention, the image acquisition module comprises a sensor configuration sub-module, a synchronous control sub-module, an image acquisition sub-module and a data integration sub-module;

the image processing module comprises an image enhancer module, an image noise reduction sub-module, a space alignment sub-module and a first feature extraction sub-module;

The three-dimensional modeling module comprises a three-dimensional reconstruction sub-module, a model gridding sub-module, a data noise reduction sub-module and a shape analysis sub-module;

the feature recognition module comprises an image preprocessing sub-module, a second feature extraction sub-module and a defect recognition sub-module;

the quality analysis module comprises a defect classification sub-module and a defect analysis sub-module;

the parameter adjustment module comprises a first parameter adjustment sub-module, a production monitoring sub-module and a strategy updating sub-module;

the optimization acquisition module comprises an anomaly analysis sub-module, an optimization scheme sub-module, a second parameter adjustment sub-module and an effect evaluation sub-module.

Compared with the prior art, the invention has the advantages and positive effects that:

in the invention, the multi-sensor synchronous acquisition technology is adopted to realize the joint acquisition of visible light, infrared and ultraviolet spectrogram images, so that the detection is more comprehensive. By using the convolutional neural network and the multi-source image fusion algorithm, the image quality is improved, the details are enhanced, and the comprehensive fusion image with higher definition is obtained. In combination with point cloud data processing, the three-dimensional model provides a more visual representation of the morphological features of the plastic master batch. With deep convolutional neural networks, defect detection and classification become more accurate, reducing the false detection rate. The application of deep reinforcement learning ensures real-time and intelligent adjustment of production parameters, enhances the capability of self-adaptive anomaly detection, and ensures continuous iterative optimization of strategies. The adoption of the reinforcement learning method provides powerful technical support for dynamic optimization of image acquisition parameters, and improves image quality and detection accuracy.

Drawings

FIG. 1 is a schematic workflow diagram of the present invention;

FIG. 2 is a S1 refinement flowchart of the present invention;

FIG. 3 is a S2 refinement flowchart of the present invention;

FIG. 4 is a S3 refinement flowchart of the present invention;

FIG. 5 is a S4 refinement flowchart of the present invention;

FIG. 6 is a S5 refinement flowchart of the present invention;

FIG. 7 is a S6 refinement flowchart of the present invention;

FIG. 8 is a system flow diagram of the present invention;

FIG. 9 is a schematic diagram of a system framework of the present invention.

Detailed Description

The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.

In the description of the present invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," and the like indicate orientations or positional relationships based on the orientation or positional relationships shown in the drawings, merely to facilitate describing the present invention and simplify the description, and do not indicate or imply that the devices or elements referred to must have a specific orientation, be configured and operated in a specific orientation, and therefore should not be construed as limiting the present invention. Furthermore, in the description of the present invention, the meaning of "a plurality" is two or more, unless explicitly defined otherwise.

Embodiment one.

Referring to fig. 1, the present invention provides a technical solution: a plastic master batch quality inspection method based on machine vision comprises the following steps:

s1: based on a plastic master batch sample, adopting a multi-sensor synchronous acquisition technology, and simultaneously acquiring visible light, infrared and ultraviolet spectrogram images to generate a multi-mode original image data set;

s2: based on a multi-mode original image dataset, performing image preprocessing by adopting a convolutional neural network, and generating a comprehensive fusion image through a multi-source image fusion algorithm;

s3: based on the comprehensive fusion image, a three-dimensional model of the plastic master batch is established by adopting a point cloud data processing method, and shape and surface analysis is carried out to generate a three-dimensional morphological analysis report;

s4: based on the comprehensive fusion image and the three-dimensional morphological analysis report, performing defect detection and classification on the plastic master batch by adopting a deep convolutional neural network to generate an automatic defect labeling diagram;

s5: based on the automatic defect labeling diagram, adopting a deep reinforcement learning algorithm to dynamically adjust production parameters, and carrying out self-adaptive anomaly detection and strategy updating to generate a quality anomaly report and an optimized detection strategy;

s6: based on the quality anomaly report, dynamically optimizing image acquisition parameters including exposure time, focal length and light source intensity by adopting a reinforcement learning method, and generating optimized image acquisition parameter settings.

The method overcomes the limitation of a single image source in recognition and detection by comprehensively utilizing the multi-mode original image dataset, and greatly improves the analysis precision and the detection reliability of the plastic master batch by fusion analysis of visible light, infrared and ultraviolet spectrograms.

By means of deep learning and application of a convolutional neural network, automatic defect detection and classification of the plastic master batch can be achieved. Through automatic defect labeling and intelligent recognition, the efficiency and the accuracy of quality detection are remarkably improved, and meanwhile, the labor intensity and the skill requirement of manual detection are reduced.

In the method, not only is the comprehensive fusion analysis of the images carried out, but also the establishment of a three-dimensional model of the plastic master batch and the comprehensive analysis of the shape and the surface of the plastic master batch are realized, which is helpful for enterprises to know the quality characteristics of products deeply and comprehensively and discover potential process defects.

The production parameters are dynamically adjusted through the deep reinforcement learning algorithm, and the self-adaptive anomaly detection and strategy updating are carried out, so that the method is not only beneficial to quickly responding to quality anomalies in the production process and reducing defective product output, but also can be carried out in the first time for optimization adjustment, and the production efficiency and the primary product yield are improved.

On the basis of quality anomaly report, the image acquisition parameters (such as exposure time, focal length, light source intensity and the like) are dynamically optimized by using a reinforcement learning method, so that the image acquisition process can be pertinently improved, and the control precision and flexibility of the whole production process can be further enhanced.

Through accurate defect identification and automatic detection flow, the method can timely and accurately find out the problems in the production process, and timely perform parameter adjustment and strategy update, so that the resource waste and the reworking cost caused by defective products are reduced.

The high-efficiency and intelligent quality inspection method not only can ensure the high quality of the product, but also can continuously improve the performance and the production efficiency of the product by continuously optimizing the production parameters, thereby enhancing the competitiveness of the enterprise product in the market and the customer satisfaction.

Referring to fig. 2, based on a plastic master batch sample, a multi-sensor synchronous acquisition technology is adopted to simultaneously acquire visible light, infrared and ultraviolet spectrogram images, and the steps for generating a multi-mode original image dataset are specifically as follows:

s101: based on the selective strategy, adopting a sensor selection algorithm to perform optimal sensor configuration and generating sensor parameter configuration;

S102: based on sensor parameter configuration, a synchronous trigger mechanism is applied, and a plurality of groups of sensors are synchronously started to obtain synchronous acquisition control signals;

s103: based on the synchronous acquisition control signal, adopting a high-speed imaging technology to acquire a visible light image, and generating a visible light original image;

s104: based on the synchronous acquisition control signal, an infrared imaging technology is utilized to acquire an infrared image, and an infrared original image is generated; s105: based on the synchronous acquisition control signal, an ultraviolet imaging technology is used for acquiring an ultraviolet image, and an ultraviolet original image is generated;

s106: based on the visible light original image, the infrared original image and the ultraviolet original image, a data integration technology is applied to combine the multi-mode images to generate a multi-mode original image data set.

Firstly, according to the characteristics and requirements of a plastic master batch sample, a selectivity strategy is adopted, and a sensor selection algorithm is applied. A specific algorithm may be a method based on an information entropy or greedy strategy to evaluate and select the optimal sensor configuration. The method is provided with a sensor set (S) and a plastic master batch sample feature set (F), and aims to find a subset (S') to maximize feature coverage:

S' = \arg \max_{S' \subseteq S} \sum_{f \in F} I(s;f），

Where (I (s; f)) represents the information gain of the sensor(s) over the feature (f). In this way, a parameter configuration of the sensor is obtained.

Next, based on the sensor parameter configuration described above, a synchronous trigger mechanism is applied to ensure that all selected sensors can operate synchronously. This can be achieved by a central control unit which, when receiving the start signal, sends control signals to all sensors for synchronous acquisition.

With the synchronization control signal, the next three steps correspond to different imaging techniques, respectively. Adopting a high-speed imaging technology to collect visible light; collecting infrared images by utilizing an infrared imaging technology; ultraviolet imaging techniques are used to acquire ultraviolet images. Among these three processes, factors that need to be considered include exposure time, sensitivity, light intensity, and the like.

The final step is to integrate the three types of images into a multi-modal raw image dataset. This dataset may be represented by a three-dimensional tensor (T), wherein each dimension corresponds to a visible, infrared and ultraviolet image, respectively. Each image can be seen as a two-dimensional matrix, examples:

t= [ i_ { visible }, i_ { infrared }, i_ { ultraviolet }, wherein (i_ { visible }, (i_ { inflroned }) and (i_ { ultraviolet }) represent visible light, infrared and ultraviolet image data, respectively.

Referring to fig. 3, based on a multi-mode original image dataset, the steps of performing image preprocessing by using a convolutional neural network and generating a comprehensive fused image by using a multi-source image fusion algorithm are specifically as follows:

s201: based on the multi-mode original image data set, adopting an image enhancement algorithm to improve the image quality and obtain an enhanced image data set;

s202: based on the enhanced image dataset, applying a median filtering algorithm to perform image noise reduction and generating a noise-reduced image dataset;

s203: based on the image dataset after noise reduction, performing spatial alignment by adopting an image registration algorithm to obtain a spatially aligned image dataset;

s204: performing brightness and contrast adjustment based on the spatially aligned image dataset by using a histogram equalization technique to generate a contrast-adjusted image dataset;

s205: based on the image data set with the adjusted contrast, extracting key features by using a SIFT feature extraction algorithm to obtain a key image feature set;

s206: based on the key image feature set, a multi-source image fusion algorithm is applied, and the multi-mode images are combined to generate a comprehensive fusion image.

S201-image enhancement:

Image enhancement is performed on a multi-modal raw image dataset, with the common enhancement algorithms being data augmentation (e.g., rotation, cropping, flipping, etc.), but image enhancement using a Generation Antagonism Network (GAN) is also contemplated. The concrete reinforcing operation is as follows:

I_{enhanced} = Augment（I_{original}）。

s202, image noise reduction:

a median filtering algorithm is applied to give an image window (such as 3x3 and 5x 5), the pixel values in the window are ordered, and the median value is taken as the value of the central pixel. The formula is:

I_{denoised}（x，y） = median{I_{enhanced}（x+i，y+j）|i，j \in W}，

where (W) is the filter window.

S203-image registration:

spatial alignment is performed using image registration algorithms such as ORB, SIFT, etc. After keypoint matching, the images are aligned using affine transformation or homography matrix:

M = Match（I_{denoised_1}， I_{denoised_2}），

I_{aligned} = Transform（I_{denoised}， M）。

s204-histogram equalization:

image brightness and contrast are adjusted using histogram equalization techniques. For each pixel, reassign the luminance value as:

I_{hist_eq}（x，y） = T（I_{aligned}（x，y）），

where (T) is a transfer function based on the cumulative histogram of the image.

S205-SIFT feature extraction: extracting key image features by using a SIFT algorithm:

K = SIFT（I_{hist_eq}），

where (K) is the keypoint and descriptor of the image.

S206-multi-source image fusion:

the multi-source image fusion algorithm involves fusion at the pixel level, feature level, or decision level. Here, consider a weighted fusion at the pixel level:

I_{fused}（x，y） = w_1 \cdot I_{source1}（x，y） + w_2 \cdot I_{source2}（x，y），

Where (w_1) and (w_2) are predefined weights.

Referring to fig. 4, based on the comprehensive fusion image, a three-dimensional model of the plastic master batch is established by adopting a point cloud data processing method, and shape and surface analysis is performed, so that a three-dimensional morphological analysis report is generated specifically by the following steps:

s301: based on the comprehensive fusion image, adopting a three-dimensional reconstruction algorithm to perform three-dimensional reconstruction to obtain a preliminary three-dimensional model;

s302: based on the preliminary three-dimensional model, performing model gridding by utilizing a three-dimensional grid generation technology to generate a gridding three-dimensional model;

s303: based on the gridding three-dimensional model, applying a point cloud filtering algorithm to perform data noise reduction to obtain a noise-reduced three-dimensional model;

s304: based on the three-dimensional model after noise reduction, adopting a surface fitting technology to analyze the shape and the surface, and generating three-dimensional shape and surface data;

s305: based on the three-dimensional shape and the surface data, a morphological operation technology is used for extracting and analyzing a specific form to obtain a form characteristic report;

s306: integrating the morphological feature report with the three-dimensional shape and surface data, and generating a three-dimensional morphological analysis report by using a data integration technology.

S301-three-dimensional reconstruction: and adopting a three-dimensional reconstruction algorithm, such as a binocular three-dimensional matching or structured light method, to reconstruct the three-dimensional of the comprehensive fusion image:

P_{3D} = StereoReconstruct（I_{fused1}， I_{fused2}），

Where (p_ {3D }) is the reconstructed point cloud data.

S302, three-dimensional grid generation: three-dimensional mesh generation techniques such as Delaunay triangulation or Maring documents algorithm are applied:

M_{3D} = MeshGeneration（P_{3D}），

s303-point cloud filtering: noise removal using statistical outlier removal or radius-based filtering:

P_{filtered} = PointCloudFilter（P_{3D}， \theta），

where "\theta" is the filtering parameter.

S304-surface fitting: shape and surface analysis is performed using surface fitting techniques, such as least squares or RANSAC algorithms:

S_{3D} = SurfaceFit（P_{filtered}， \phi），

wherein, (\phi) is the fitting parameter.

S305-morphological operation: morphological operation techniques such as erosion, dilation, open and close operations are applied to extract and analyze specific morphological features:

F_{morph} = MorphologyOperation（S_{3D}， \kappa），

wherein, (\kappa) is the morphologically manipulated core.

S306-generating a three-dimensional morphological analysis report: combining the morphology feature report with the three-dimensional shape and surface data, integrating all the data to generate a final analysis report:

R_{3D} = IntegrateReport（F_{morph}， S_{3D}），

referring to fig. 5, based on the comprehensive fusion image and the three-dimensional morphological analysis report, the method adopts the deep convolutional neural network to detect and classify the defects of the plastic master batch, and the steps for generating the automatic defect labeling chart specifically include:

s401: based on the comprehensive fusion image, adopting an image enhancement algorithm to perform image preprocessing to generate an enhanced plastic master batch image;

S402: based on the reinforced plastic master batch image, performing feature extraction by adopting a SIFT technology to obtain a plastic master batch feature set;

s403: based on the plastic master batch characteristic set, performing defect identification by adopting a deep convolutional neural network to obtain a defect identification label;

s404: and combining the defect identification label and the reinforced plastic master batch image, and adopting a classification algorithm to classify the defects to generate an automatic defect labeling chart.

S401-image enhancement: the comprehensive fused image is preprocessed using common image enhancement algorithms, such as histogram equalization: i_ { enhanced } = histogramequation (i_ { fused }).

S402, feature extraction: extracting key points and feature descriptors from the enhanced image by adopting a Scale Invariant Feature Transform (SIFT) algorithm:

（K， D） = SIFT（I_{enhanced}），

wherein (K) represents a key point and (D) represents a descriptor.

S403, defect identification:

and performing defect identification on the SIFT feature descriptor by using a Deep Convolutional Neural Network (DCNN). Consider a pre-trained DCNN model as (m_ { DCNN }):

L = M_{dcnn}（D），

wherein, (L) is a defect identification tag.

S404, defect classification and marking:

based on the defect identification tag and the enhanced image, a classification algorithm such as a Support Vector Machine (SVM) is applied:

T_{defect} = SVM（L， I_{enhanced}），

And then carrying out defect marking on the enhanced image:

I_{annotated} = Annotate（I_{enhanced}， T_{defect}），

referring to fig. 6, based on the automatic defect labeling chart, the method adopts a deep reinforcement learning algorithm to dynamically adjust production parameters, and performs self-adaptive anomaly detection and strategy update, and the steps of generating a quality anomaly report and an optimized detection strategy are specifically as follows:

s501: analyzing defect distribution by adopting a data analysis technology based on the automatic defect labeling diagram to obtain a defect analysis report;

s502: based on the defect analysis report, adopting a deep reinforcement learning algorithm to adjust production parameters and generating adjusted production parameters;

s503: monitoring the production process by using the adjusted production parameters and adopting a self-adaptive abnormality detection technology to generate an abnormality detection report;

s504: and optimizing the detection strategy by adopting a strategy updating algorithm according to the abnormal detection report to obtain a quality abnormal report.

S501-defect analysis: first, data analysis is performed based on an automatic defect labeling diagram:

r_ { defect } = data analysis (i_ { annotation }),

wherein, (R_ { defect }) represents a defect analysis report.

S502-production parameter adjustment:

and adjusting production parameters by using a deep reinforcement learning model. Assuming that the state is the current production environment and the action is the adjustment parameter, the rewards are based on the degree of defect reduction:

P_{adjusted} = DRLModel（R_{defect}， S_{current}），

Where (P_ { adjusted }) is the adjusted production parameter and (S_ { current }) is the current production state.

S503-anomaly detection: using the adjusted production parameters, applying an adaptive anomaly detection technique:

R_{anomaly} = AnomalyDetection（P_{adjusted}， S_{current}），

wherein, (R_ { analog }) is an anomaly detection report.

S504, strategy updating and exception reporting: updating the detection strategy according to the abnormal detection report:

R_{quality}， S_{update} = PolicyUpdate（R_{anomaly}， P_{adjusted}），

wherein, (R_ { quality }) is a quality anomaly report, (S_ { update }) is an optimized detection strategy.

Referring to fig. 7, based on the quality anomaly report, the image acquisition parameters including exposure time, focal length, and light source intensity are dynamically optimized by using a reinforcement learning method, and the steps for generating the optimized image acquisition parameter settings are specifically as follows:

s601: according to the quality abnormality report, adopting an analysis technology to discuss the abnormality cause and obtain an abnormality cause analysis report;

s602: based on the analysis report of the abnormal reasons, applying a reinforcement learning method, and providing an image acquisition optimization scheme to generate a parameter optimization scheme;

s603: combining a parameter optimization scheme, adjusting exposure time, focal length and light source intensity, and generating primarily optimized image acquisition parameters;

s604: the image acquisition is carried out again by utilizing the preliminarily optimized image acquisition parameters, the effect is estimated, and an effect estimation report is generated;

S605: and (3) carrying out parameter refinement and perfecting according to the effect evaluation report, and generating optimized image acquisition parameter setting.

S601-analysis of abnormality cause: first, analysis is performed on a quality anomaly report:

r_ { cause } = anomaly analysis (r_ { quality }),

wherein (R_ { reliability }) represents an abnormality cause analysis report and (R_ { quality }) is a quality abnormality report.

S602-parameter optimization scheme: analyzing the abnormal cause report by using a reinforcement learning model to propose an image acquisition optimization scheme:

s_ { scheme } = RLModel (r_ { cause }),

wherein, (S_ { summation }) is a parameter optimization scheme.

S603-preliminary parameter adjustment: according to the parameter optimization scheme, adjusting image acquisition parameters:

P_{initial} = AdjustParameters（S_{suggestion}），

where (p_ { initial }) is the initially optimized image acquisition parameter.

S604-effect evaluation: and (3) re-carrying out image acquisition by utilizing the primarily optimized image acquisition parameters, and evaluating the effect:

R_{evaluation} = ImageEvaluation（P_{initial}），

where (R_ { evaluation }) is the effect evaluation report.

S605-parameter refinement and perfecting: according to the effect evaluation report, refining and perfecting image acquisition parameters:

P_{optimized} = RefineParameters（R_{evaluation}），

wherein, (P_ { optimized }) is the optimized image acquisition parameter setting.

Referring to fig. 8, a plastic master batch quality inspection system based on machine vision is used for executing the plastic master batch quality inspection method based on machine vision, and the system comprises an image acquisition module, an image processing module, a three-dimensional modeling module, a feature recognition module, a quality analysis module, a parameter adjustment module and an optimization acquisition module.

The image acquisition module acquires visible light, infrared and ultraviolet images by adopting a synchronous trigger mechanism based on a sensor selection algorithm to generate a multi-mode original image data set;

the image processing module generates an image dataset with improved quality by adopting an image enhancement algorithm, a median filtering algorithm, an image registration algorithm and a histogram equalization technology based on the multi-mode original image dataset, and finally generates a comprehensive fusion image;

the three-dimensional modeling module generates a meshed three-dimensional model and three-dimensional shape surface data by utilizing a three-dimensional reconstruction algorithm and a three-dimensional mesh generation technology based on the comprehensive fusion image;

the feature recognition module is used for preprocessing the comprehensive fusion image, extracting features, performing defect recognition by using a deep convolutional neural network, and generating a defect recognition tag;

The quality analysis module classifies defects based on the defect identification labels, generates an automatic defect labeling chart, analyzes defect distribution and generates a defect analysis report;

the parameter adjustment module adjusts production parameters based on the defect analysis report by using a deep reinforcement learning algorithm, generates adjusted production parameters, monitors the production process, generates an abnormal detection report, optimizes the detection strategy by a strategy updating algorithm, and finally generates a quality abnormal report;

the optimization acquisition module analyzes the abnormal reasons, proposes an optimization scheme, generates a parameter optimization scheme, adjusts image acquisition parameters, generates primarily optimized image acquisition parameter settings, re-acquires images, evaluates effects and finally obtains the optimized image acquisition parameter settings.

Firstly, the multi-mode image acquisition function of the system enables the characteristics of the plastic master batch to be observed under a plurality of spectrums, so that accurate and reliable data information can be obtained under different environments and physical conditions. By capturing visible, infrared and ultraviolet images, nuances and potential defects of the material can be captured from multiple dimensions, enhancing the accuracy and reliability of the detection.

And secondly, various image processing technologies such as an image enhancement algorithm, a median filtering algorithm and the like are introduced, so that the quality and resolution of the image are obviously improved, the subsequent feature identification and quality analysis have a clearer and more accurate data base, and the possibility of false detection and omission is greatly reduced.

Meanwhile, by implementing a three-dimensional modeling technology, the system can more accurately capture a plurality of physical characteristics such as the shape, the size, the surface texture and the like of the plastic master batch, and provides omnibearing data support for accurate quality detection. In particular, three-dimensional modeling techniques have shown their irreplaceable importance for spatial defects and deformations that are not perceptible in two-dimensional images.

In the aspect of feature recognition, the introduction of the deep convolutional neural network greatly widens the recognition capability of complex and tiny defects. Compared with the traditional method, the deep learning brings higher level of automation and intelligence, greatly lightens the labor burden and also remarkably improves the detection efficiency and the detection precision.

From quality analysis and parameter adjustment aspect, the use of degree of depth reinforcement study makes the system not only can carry out accurate quality detection and analysis based on current data, can also through constantly learning, intelligent optimization production parameter and detection strategy, ensure production process's stability to in time send the early warning when meetting the problem, thereby prevented the quality accident that probably takes place.

Referring to fig. 9, the image acquisition module includes a sensor configuration sub-module, a synchronization control sub-module, an image acquisition sub-module, and a data integration sub-module; the image processing module comprises an image enhancer module, an image noise reduction sub-module, a space alignment sub-module and a first characteristic extraction sub-module;

The three-dimensional modeling module comprises a three-dimensional reconstruction sub-module, a model gridding sub-module, a data noise reduction sub-module and a shape analysis sub-module;

the feature recognition module comprises an image preprocessing sub-module, a second feature extraction sub-module and a defect recognition sub-module;

the quality analysis module comprises a defect classification sub-module and a defect analysis sub-module;

the parameter adjustment module comprises a first parameter adjustment sub-module, a production monitoring sub-module and a strategy updating sub-module;

the optimization acquisition module comprises an anomaly analysis sub-module, an optimization scheme sub-module, a second parameter adjustment sub-module and an effect evaluation sub-module.

And an image acquisition module:

sensor configuration submodule: by optimizing the layout and performance configuration of the sensor, higher resolution and quality image data can be obtained.

And a synchronization control submodule: data acquisition synchronization between different sensors and devices is ensured to avoid data inconsistencies.

An image acquisition sub-module: is responsible for acquiring the original image data and provides an initial source of data.

A data integration sub-module: the data from the different sensors and devices are integrated together to provide a complete data set for subsequent processing.

An image processing module:

An image enhancer module: by enhancing contrast, brightness, etc., the visual quality of the image can be improved, contributing to a clearer analysis of the image.

An image noise reduction sub-module: noise in the image is removed, and the image quality and accuracy are improved.

Spatial alignment submodule: the images of different viewing angles or sensors are accurately spatially aligned to provide accurate input for three-dimensional modeling.

A first feature extraction sub-module: and extracting key features, and facilitating subsequent processing and analysis.

And a three-dimensional modeling module:

three-dimensional reconstruction sub-module: the image data shot from a plurality of angles are converted into a three-dimensional model, so that three-dimensional restoration of the target object is realized.

Model meshing submodule: the three-dimensional model is converted into grid or point cloud data, which provides a data format for further analysis.

And a data noise reduction sub-module: noise in the three-dimensional data is removed, and the precision of the model is improved.

Shape analysis submodule: the geometry of the model is analyzed to provide information about the target object.

And the characteristic recognition module is used for:

an image preprocessing sub-module: the input image is further processed in preparation for feature extraction.

A second feature extraction sub-module: higher level features, such as texture, shape, etc., are extracted from the image.

Defect identification sub-module: detecting and identifying defects or anomalies in the image facilitates quality control.

And a mass analysis module:

defect classification sub-module: the detected defects are classified to determine their severity.

Defect analysis sub-module: the nature and location of the defect is analyzed, providing more detailed information for improving the process.

Parameter adjustment module:

a first parameter adjustment sub-module: system parameters are adjusted to optimize the image acquisition and processing process.

Production monitoring submodule: the production process is monitored in real time, problems are found in time, and measures are taken.

Policy update sub-module: and updating the system strategy according to the monitoring result to improve the performance.

And (3) an optimization acquisition module:

an anomaly analysis sub-module: abnormal conditions, such as equipment failure or quality problems, are identified to take appropriate action.

Optimization scheme sub-module: an improvement scheme is provided to improve the production efficiency and quality.

And a second parameter adjustment sub-module: the parameters are further adjusted to achieve better performance.

An effect evaluation sub-module: the system effect is evaluated to verify the effect of the improvement.

The present invention is not limited to the above embodiments, and any equivalent embodiments which can be changed or modified by the technical disclosure described above can be applied to other fields, but any simple modification, equivalent changes and modification made to the above embodiments according to the technical matter of the present invention will still fall within the scope of the technical disclosure.

Claims (10)

1. The plastic master batch quality inspection method based on machine vision is characterized by comprising the following steps of:

based on a plastic master batch sample, adopting a multi-sensor synchronous acquisition technology, and simultaneously acquiring visible light, infrared and ultraviolet spectrogram images to generate a multi-mode original image data set;

based on the multi-mode original image data set, performing image preprocessing by adopting a convolutional neural network, and generating a comprehensive fusion image through a multi-source image fusion algorithm;

based on the comprehensive fusion image, a three-dimensional model of the plastic master batch is established by adopting a point cloud data processing method, and shape and surface analysis is carried out to generate a three-dimensional morphological analysis report;

based on the comprehensive fusion image and the three-dimensional morphological analysis report, performing defect detection and classification on the plastic master batch by adopting a deep convolutional neural network to generate an automatic defect labeling chart;

based on the automatic defect labeling diagram, adopting a deep reinforcement learning algorithm to dynamically adjust production parameters, and carrying out self-adaptive anomaly detection and strategy updating to generate a quality anomaly report and an optimized detection strategy;

and dynamically optimizing image acquisition parameters, including exposure time, focal length and light source intensity, by adopting a reinforcement learning method based on the quality anomaly report, and generating optimized image acquisition parameter settings.

2. The machine vision-based plastic master batch quality inspection method according to claim 1, wherein the step of simultaneously acquiring visible light, infrared and ultraviolet spectrogram images and generating a multi-mode original image data set is specifically as follows:

based on the selective strategy, adopting a sensor selection algorithm to perform optimal sensor configuration and generating sensor parameter configuration;

based on the sensor parameter configuration, a synchronous trigger mechanism is applied to synchronously start a plurality of groups of sensors to obtain synchronous acquisition control signals;

based on the synchronous acquisition control signal, adopting a high-speed imaging technology to acquire a visible light image, and generating a visible light original image;

based on the synchronous acquisition control signal, an infrared imaging technology is utilized to acquire an infrared image, and an infrared original image is generated;

based on the synchronous acquisition control signal, an ultraviolet imaging technology is used for acquiring an ultraviolet image, and an ultraviolet original image is generated;

based on the visible light original image, the infrared original image and the ultraviolet original image, a data integration technology is applied, and the multi-mode images are combined to generate a multi-mode original image data set.

3. The machine vision-based plastic master batch quality inspection method according to claim 1, wherein the steps of performing image preprocessing by using a convolutional neural network based on the multi-mode original image dataset and generating a comprehensive fusion image by using a multi-source image fusion algorithm are specifically as follows:

based on the multi-mode original image data set, adopting an image enhancement algorithm to improve the image quality and obtaining an enhanced image data set;

based on the enhanced image dataset, applying a median filtering algorithm to perform image noise reduction and generating a noise-reduced image dataset;

based on the noise-reduced image data set, performing spatial alignment by adopting an image registration algorithm to obtain a spatially aligned image data set;

performing brightness and contrast adjustment based on the spatially aligned image dataset by using a histogram equalization technique to generate a contrast-adjusted image dataset;

based on the image dataset with the adjusted contrast, extracting key features by using a SIFT feature extraction algorithm to obtain a key image feature set;

and based on the key image feature set, applying a multi-source image fusion algorithm, and combining the multi-mode images to generate a comprehensive fusion image.

4. The machine vision-based plastic master batch quality inspection method according to claim 1, wherein the step of creating a three-dimensional model of the plastic master batch and performing shape and surface analysis based on the comprehensive fusion image by using a point cloud data processing method to generate a three-dimensional morphological analysis report is specifically as follows:

based on the comprehensive fusion image, adopting a three-dimensional reconstruction algorithm to perform three-dimensional reconstruction to obtain a preliminary three-dimensional model;

based on the preliminary three-dimensional model, performing model gridding by utilizing a three-dimensional grid generation technology to generate a gridding three-dimensional model;

based on the meshed three-dimensional model, applying a point cloud filtering algorithm to perform data noise reduction to obtain a noise-reduced three-dimensional model;

based on the three-dimensional model after noise reduction, adopting a surface fitting technology to analyze the shape and the surface, and generating three-dimensional shape and surface data;

based on the three-dimensional shape and the surface data, extracting and analyzing a specific form by using a morphological operation technology to obtain a form characteristic report;

integrating the morphological feature report with the three-dimensional shape and surface data, and generating a three-dimensional morphological analysis report by using a data integration technology.

5. The machine vision-based plastic master batch quality inspection method according to claim 1, wherein the step of performing defect detection and classification on plastic master batches by using a deep convolutional neural network based on the comprehensive fusion image and the three-dimensional morphological analysis report to generate an automatic defect labeling chart is specifically as follows:

based on the comprehensive fusion image, adopting an image enhancement algorithm to perform image preprocessing to generate an enhanced plastic master batch image;

based on the reinforced plastic master batch image, performing feature extraction by adopting a SIFT technology to obtain a plastic master batch feature set;

based on the plastic master batch characteristic set, performing defect identification by adopting a deep convolutional neural network to obtain a defect identification tag;

and combining the defect identification label and the reinforced plastic master batch image, and adopting a classification algorithm to classify defects to generate an automatic defect labeling chart.

6. The machine vision-based plastic master batch quality inspection method according to claim 1, wherein the steps of dynamically adjusting production parameters by adopting a deep reinforcement learning algorithm based on the automatic defect labeling diagram, performing self-adaptive anomaly detection and policy updating, and generating a quality anomaly report and an optimized detection policy are specifically as follows:

Analyzing defect distribution by adopting a data analysis technology based on the automatic defect labeling diagram to obtain a defect analysis report;

based on the defect analysis report, adopting a deep reinforcement learning algorithm to adjust production parameters and generating adjusted production parameters;

monitoring a production process by using the adjusted production parameters and adopting a self-adaptive abnormality detection technology to generate an abnormality detection report;

and optimizing the detection strategy by adopting a strategy updating algorithm according to the abnormal detection report to obtain a quality abnormal report.

7. The machine vision-based plastic master batch quality inspection method according to claim 1, wherein the step of dynamically optimizing image acquisition parameters including exposure time, focal length, light source intensity by using reinforcement learning method based on the quality anomaly report, and generating optimized image acquisition parameter settings specifically comprises the steps of:

according to the quality abnormality report, adopting an analysis technology to discuss the abnormality cause and obtain an abnormality cause analysis report;

based on the abnormal cause analysis report, applying a reinforcement learning method, and providing an image acquisition optimization scheme to generate a parameter optimization scheme;

combining the parameter optimization scheme, adjusting exposure time, focal length and light source intensity, and generating primarily optimized image acquisition parameters;

Carrying out image acquisition again by utilizing the primarily optimized image acquisition parameters, evaluating the effect of the image acquisition parameters and generating an effect evaluation report;

and according to the effect evaluation report, carrying out parameter refinement and perfecting to generate optimized image acquisition parameter setting.

8. The plastic master batch quality inspection system based on machine vision is characterized by being used for executing the plastic master batch quality inspection method based on machine vision as set forth in any one of claims 1-7, and comprises an image acquisition module, an image processing module, a three-dimensional modeling module, a characteristic recognition module, a quality analysis module, a parameter adjustment module and an optimization acquisition module.

9. The machine vision-based plastic master batch quality inspection system according to claim 8, wherein the image acquisition module acquires visible light, infrared and ultraviolet images by adopting a synchronous trigger mechanism based on a sensor selection algorithm to generate a multi-mode original image dataset;

the image processing module generates an image dataset with improved quality by adopting an image enhancement algorithm, a median filtering algorithm, an image registration algorithm and a histogram equalization technology based on the multi-mode original image dataset, and finally generates a comprehensive fusion image;

The three-dimensional modeling module generates a meshed three-dimensional model and three-dimensional shape surface data by utilizing a three-dimensional reconstruction algorithm and a three-dimensional mesh generation technology based on the comprehensive fusion image;

the feature recognition module is used for preprocessing the comprehensive fusion image, extracting features, performing defect recognition by using a deep convolutional neural network, and generating a defect recognition tag;

the quality analysis module classifies defects based on the defect identification labels, generates an automatic defect labeling chart, analyzes defect distribution and generates a defect analysis report;

the parameter adjustment module adjusts production parameters based on the defect analysis report by using a deep reinforcement learning algorithm, generates adjusted production parameters, monitors the production process, generates an abnormal detection report, optimizes the detection strategy by a strategy updating algorithm, and finally generates a quality abnormal report;

the optimization acquisition module analyzes the abnormal reasons, proposes an optimization scheme, generates a parameter optimization scheme, adjusts image acquisition parameters, generates primarily optimized image acquisition parameter settings, re-acquires images, evaluates effects and finally obtains the optimized image acquisition parameter settings.

10. The machine vision-based plastic master batch quality inspection system of claim 8, wherein the image acquisition module comprises a sensor configuration sub-module, a synchronization control sub-module, an image acquisition sub-module, and a data integration sub-module;

The image processing module comprises an image enhancer module, an image noise reduction sub-module, a space alignment sub-module and a first feature extraction sub-module;

the three-dimensional modeling module comprises a three-dimensional reconstruction sub-module, a model gridding sub-module, a data noise reduction sub-module and a shape analysis sub-module;

the feature recognition module comprises an image preprocessing sub-module, a second feature extraction sub-module and a defect recognition sub-module;

the quality analysis module comprises a defect classification sub-module and a defect analysis sub-module;

the parameter adjustment module comprises a first parameter adjustment sub-module, a production monitoring sub-module and a strategy updating sub-module;

the optimization acquisition module comprises an anomaly analysis sub-module, an optimization scheme sub-module, a second parameter adjustment sub-module and an effect evaluation sub-module.