CN113538342B - A method for detecting the coating quality of aluminum aerosol cans based on convolutional neural network - Google Patents

Abstract

The invention discloses a convolutional neural network-based aluminum aerosol can coating quality detection method, which comprises the following steps: obtaining a photographing sample of an aluminum aerosol can product on a production line; classifying the photographed samples to finish marking the original image data; performing data enhancement and constructing a sample data set; establishing a grid model: the method comprises the steps of training a Fine tuning-GoogLeNet and a Fine tuning-ResNet18, inputting new image data, carrying out image classification prediction by using a trained defect discrimination model, and taking an image data label with highest prediction probability as a judgment basis. By adopting the invention, the quality detection accuracy of the inner coating and the outer coating of the aluminum aerosol can coating respectively reaches 99.38 percent and 96.4 percent, the efficiency is high, and compared with the manual visual detection, the speed is faster, the accuracy and the reliability are higher; compared with the traditional machine vision detection mode based on feature extraction, the method has stronger scene adaptation capability and better robustness, and can reduce the enterprise cost and improve the efficiency of product quality detection.

Description

A method for detecting the coating quality of aluminum aerosol cans based on convolutional neural network

technical field

The invention relates to the technical field of product quality detection, in particular to a method for detecting the coating quality of an aluminum aerosol can based on a convolutional neural network.

Background technique

Aluminum aerosol cans are favored by the packaging market due to their light weight and strong plasticity, and are widely used in industrial products, personal protection and cosmetic products. During the production process, coatings are sprayed and printed on the inner and outer surfaces of the tank to shield the contents of the tank and foreign substances from direct contact with aluminum. Therefore, its spraying and printing process directly affects the quality, qualification rate and production efficiency of the product. However, due to the influence of many uncertain factors in the spraying and printing process, it is easy to produce defects such as coating bubbles, stains, and unevenness of the tank body, which makes the product not only unsightly, but also affects the use (the unevenness of the tank body will affect the pressure bearing capacity of the product, resulting in The service life is shortened, and even affects the personal safety of users), so it is necessary to detect the coating quality of aluminum aerosol cans.

Traditional aluminum aerosol can coating quality inspection methods currently mainly rely on human visual inspection, which has disadvantages such as low detection efficiency, high labor costs, low accuracy, and unreliable detection results. As a rapidly developing branch of artificial intelligence, machine vision technology has been applied in many automated inspection fields, but most of them are applied to the detection of steel, aluminum cans and flat printed matter based on traditional image feature recognition methods. Convolutional neural network based machine vision inspection system for aerosol can coating quality.

Contents of the invention

The technical problem to be solved by the embodiments of the present invention is to provide a method for detecting the coating quality of an aluminum aerosol can based on a convolutional neural network. It can realize simultaneous detection of defects such as air bubbles, bumps, and stains in the inner and outer coatings of aluminum aerosol cans.

In order to solve the above technical problems, an embodiment of the present invention provides a method for detecting the coating quality of an aluminum aerosol can based on a convolutional neural network, including the following steps:

S1: Obtain photographic samples of aluminum aerosol cans on the production line;

S2: Divide the photographed samples into two types: inner surface images and outer surface images, and classify them respectively to complete the original image data labeling;

S3: Perform data enhancement and construct a sample data set of inner and outer coatings of aluminum aerosol cans;

S4: adjust image size;

S5: Based on the method of transfer learning, fine-tune the GoogLeNet and ResNet18 models to obtain two convolutional neural network models suitable for the image data set of aluminum can cold spray coating: Fine tuning-GoogLeNet and Finetuning-ResNet18;

S6: Adjust the initial training parameters of the network, use the aluminum aerosol can inner coating data set to perform model training on the Fine tuning-GoogLeNet, and use the aluminum aerosol can outer coating data set to perform model training on Fine tuning-ResNet18, Obtain a trained aluminum aerosol can defect discrimination model based on Fine tuning-GoogLeNet and Fine tuning-ResNet18;

S7: Input new image data, use the trained defect discrimination model to perform image classification prediction, and use the image data label with the highest prediction probability as the judgment basis.

Among them, the Fine tuning-GoogLeNet is to use the image data of the inner coating of the cold spraying of the aluminum can, and adjust it on the basis of the GoogLeNet InceptionV1 network model. The adjustment method includes:

S11: freeze other layers, so that the input layer accepts an input image size of 227*227*1;

S12: Replace the local response normalization layer in the original network with a batch normalization layer;

S13: Set Dropout in front of the fully connected layer;

S14: The output nodes of the fully connected layer are set to 3, and the Softmax function is used after the fully connected layer to realize image classification output.

Wherein, the method of the batch normalization layer includes:

S121: For the input minimum batch data Define the normalized network response as ;

S122: Calculate the mean value of each minimum batch of data: ;

S123: Calculate the variance of each minimum batch of data: ;

S124: Use the obtained mean value and variance to perform normalization processing on the minimum batch data, so that the mean value is 0 and the variance is 1:

S125: Introduce trainable parameters , to zoom and pan the data: /> .

Wherein, the Fine tuning-ResNet18 is adjusted in the ResNet18 network model, including the following steps:

1) Freeze other layers so that the input layer accepts an input image size of 224*224*1;

2) Set Dropout in front of the fully connected layer;

3) The output nodes of the FC layer are set to 5, and the Softmax function is used behind it to realize the image classification output.

Wherein, the step S7 further includes: if it is a normal image label, recording the current product as a normal product and simultaneously recording it as an inspected product; otherwise, recording it as a defective product corresponding to the label and simultaneously recording it as an inspected product.

The implementation of the embodiment of the present invention has the following beneficial effects: the present invention builds a defect discrimination model for aluminum aerosol cans, and realizes the detection of product quality. In the established data set, the accuracy of the quality detection of the inner and outer coatings reaches 99.38% respectively. % and 96.4%; inner and outer coating single-can inspection speed is faster, faster, more accurate and more reliable than manual visual inspection; scene adaptability is better than traditional machine vision inspection methods based on feature extraction Stronger and more robust, it can reduce enterprise costs and improve the efficiency of product quality inspection.

Description of drawings

Figure 1 is a schematic diagram of the construction process of the aluminum aerosol can image data set;

Fig. 2 is a schematic diagram of the training process of the defect discrimination model of the inner and outer coatings of aluminum aerosol cans;

Fig. 3 is a schematic flow chart of the method for detecting the quality of the inner and outer coatings of an aluminum aerosol can;

Figure 4 is a schematic diagram of the Fine tuning-GoogLeNet network structure design;

Figure 5 is a schematic diagram of the structure design of Fine tuning-ResNet18;

Fig. 6 is the inspection result of the inner coating quality of an aluminum aerosol can with bubble defects in the inner coating;

Figure 7 shows the quality inspection results of the outer coating of an aluminum aerosol can with the defect of the outer coating concavity.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings.

A method for detecting the coating quality of an aluminum aerosol can based on a convolutional neural network in an embodiment of the present invention is implemented through the following implementation steps.

1. Collect aluminum aerosol cans on the production line, including normal products and defective products, as shown in Figure 1.

2. Use a black and white camera to take pictures of the collected products on their inner and outer surfaces. According to the obtained samples, they are divided into two categories: inner surface images and outer surface images, and each category is divided into several subcategories according to normal products and products with different defects, and the original image data marking is completed.

3. Carry out data enhancement on the original image and construct a sample data set of inner and outer coatings of aluminum aerosol cans. Using spatial scale transformation (horizontal flip, vertical flip, horizontal-vertical flip) and adding noise (Gaussian blur processing) to expand the original data and enhance the fault tolerance and robustness of the subsequent convolutional neural network model. Based on the enhanced image data, a sample data set of the inner and outer coatings of aluminum aerosol cans is constructed. The data set divides the normal image and each type of coating image with defects into several categories, and labels each type of image.

The Gaussian blur processing method used in it is based on the principle of two-dimensional Gaussian distribution:

to the original image , where /> Represents the coordinate points of the original image, using the Gaussian convolution kernel constructed by the two-dimensional Gaussian function, in the original image /> Sliding window convolution is performed on , and the process can be expressed as , get /> It is the image processed by Gaussian blur.

4. Resize the image. The image pixel size of the inner coating data is adjusted to [227,227], and the pixel size of the outer coating data image is adjusted to [224,224].

5. Build a classification convolutional neural network. Based on the transfer learning method, the GoogLeNet and ResNet18 models were fine-tuned to obtain two convolutional neural network models suitable for the aluminum can cold spray coating image data set: Fine tuning-GoogLeNet and Fine tuning-ResNet18, as shown in Figures 2 and 4 , 5 shown.

The network structure of Fine tuning-GoogLeNet is adjusted based on the GoogLeNet InceptionV1 network model for the image data of the inner coating of aluminum can cold spraying. The adjustment method is as follows:

1) Freeze other layers so that the input layer accepts an input image size of 227*227*1;

2) Replace the local response normalization layer (Local Response Normalization, LRN) in the original network with the batch normalization layer (Batch Normalization, BN);

Among them, the algorithm flow of BN layer is as follows:

The minimum batch of input data (mini-batch) , define the normalized network response as /> .

a) First calculate the mean of each mini-batch: ;

b) Calculate the variance of each mini-batch: ;

c) Use the obtained mean and variance to normalize the mini-batch so that the mean is 0 and the variance is 1:

d) ;

e) Introduce trainable parameters , to zoom and pan the data: />

3) Set Dropout (temporarily hidden neurons) in front of the fully connected layer;

4) The output nodes of the fully connected layer (FC) are set to 3, and the Softmax function is used after the fully connected layer to realize image classification output.

The network can be trained and classified on three types of images of inner coatings (normal, air bubbles, raised inner surface).

The Fine tuning-ResNet18 network structure is adjusted based on the ResNet18 network model for the aluminum can cold spray coating data set. The method is as follows:

1) Freeze the other layers so that the input layer accepts an input image of size ;

2) Set Dropout in front of the fully connected layer;

3) The output nodes of the FC layer are set to 5, and the Softmax function is used behind it to realize the image classification output.

The network can be trained and classified on five categories of images of the exterior coating (normal, air bubbles, tank dents, cuff dents, and stains).

6. Network model training. Adjust the initial training parameters of the network, use Fine tuning-GoogLeNet and Fine tuning-ResNet18 to perform model training on the inner coating and outer coating data sets of aluminum aerosol cans, and obtain the trained model based on Fine tuning-GoogLeNet and Fine tuning-ResNet18 Defect discrimination model for aluminum aerosol cans.

7. Based on the trained defect discrimination model, detection and judgment are carried out.

8. Input new image data, use the trained defect discrimination model for image classification prediction, and use the image data label with the highest prediction probability as the basis for judgment. If it is a normal image label, record the current product as a normal product and record it as a detected product at the same time; otherwise, record it as a defective product corresponding to the label and record it as a detected product at the same time, as shown in Figure 3.

Applying the method of the present invention, the result of detecting bubble defects in the inner coating of aluminum aerosol cans is as shown in Figure 6, and detecting the quality of the outer coating of aluminum aerosol cans with the concavity of the outer coating The result of the defect is shown in Figure 7.

Implement the present invention, have the following advantages:

1. The automatic detection method for the spraying and printing process characteristics of aluminum aerosol cans is an effective supplement to the product quality inspection of the existing high-speed aluminum aerosol can production line;

2. This method is different from human eye detection and traditional machine vision based on feature extraction to detect product quality. It uses a method based on convolutional neural network and transfer learning, using Fine tuning-GoogLeNet and Fine tuning-ResNet18 networks , build a defect discrimination model for aluminum aerosol cans, and realize the detection of product quality. In the established data set, the accuracy rates of inner and outer coating quality inspections reached 99.38% and 96.4% respectively; the single-can inspection speed of inner and outer coatings is faster. Compared with manual visual detection, it has faster speed, higher accuracy and reliability; it has stronger scene adaptability and better robustness than traditional feature extraction-based machine vision detection methods. It can reduce the cost of the enterprise and improve the efficiency of product quality inspection.

The above disclosure is only a preferred embodiment of the present invention, which certainly cannot limit the scope of rights of the present invention. Therefore, equivalent changes made according to the claims of the present invention still fall within the scope of the present invention.

Claims (5)

1. The method for detecting the quality of the aluminum aerosol can coating based on the convolutional neural network is characterized by comprising the following steps of:

s1: obtaining a photographing sample of an aluminum aerosol can product on a production line;

s2: dividing the shot sample into two major categories of an inner surface image and an outer surface image, and respectively classifying to finish marking of original image data;

s3: carrying out data enhancement and constructing an aluminum aerosol can inner and outer coating sample data set;

s4: adjusting the size of the image;

s5: based on a transfer learning mode, fine tuning is carried out on GoogLeNet and ResNet18 models to obtain two convolutional neural network models suitable for aluminum can cold spray coating image data sets: fine tuning-google net and Fine tuning-ResNet18;

s6: adjusting initial training parameters of a network, performing model training on the Fine tuning-GoogLeNet by using an inner coating data set of an aluminum aerosol can, and performing model training on the Fine tuning-ResNet18 by using an outer coating data set of the aluminum aerosol can to obtain a trained aluminum aerosol can defect judging model based on the Fine tuning-GoogLeNet and the Fine tuning-ResNet18;

s7: inputting new image data, carrying out image classification prediction by using a trained defect discrimination model, and taking an image data label with highest prediction probability as a judgment basis.

2. The convolutional neural network-based aluminum aerosol can coating quality detection method of claim 1, wherein the Fine tuning-google net is based on GoogLeNet InceptionV network model for using aluminum can cold spray inner coating image data, the adjustment method comprises:

s11: freezing other layers to make the input layer accept the input image size of 227 x 1;

s12: replacing the local response normalization layer in the original network with a batch normalization layer;

s13: a Dropout is arranged in front of the full connection layer;

s14: the number of output nodes of the full-connection layer is 3, and after the full-connection layer, the Softmax function is used for realizing the classified output of the images.

3. The convolutional neural network-based aluminum aerosol can coating quality detection method of claim 2, wherein the batch normalization layer method comprises:

s121: for minimum batch data inputDefining normalized network responses as；

S122: calculating the mean value of each minimum batch of data:；

s123: calculating the variance of each minimum batch of data:；

s124: and carrying out normalization processing on the minimum batch data by using the obtained mean value and variance to ensure that the mean value is 0 and the variance is 1:

s125: introducing trainable parametersScaling and translating the data: />。

4. The convolutional neural network-based aluminum aerosol can coating quality detection method of claim 1, wherein the Fine tuning-ResNet18 is tuned in a ResNet18 network model, comprising the steps of:

1) Freezing other layers to make the input layer accept the input image with 224 x 1;

2) A Dropout is arranged in front of the full connection layer;

3) The FC layer output nodes are set to 5, and then the Softmax function is used to realize image classification output.

5. The method for detecting the coating quality of an aluminum aerosol can based on a convolutional neural network according to any one of claims 1 to 4, wherein the step S7 further comprises: if the current product is the normal image label, recording the current product as a normal product and recording the current product as a detected product; otherwise, recording the product as the corresponding defect product of the label, and recording the product as the detected product.