CN115290596A - FCN-ACGAN data enhancement-based hidden dangerous goods identification method and equipment - Google Patents

Abstract

The invention discloses a hidden dangerous goods identification method and equipment based on FCN-ACGAN data enhancement, relating to the technical field of goods detection and comprising the following steps: the method comprises the steps of preprocessing real terahertz spectrum data of an article acquired in advance to obtain a real sample, generating a simulation sample by using a pre-trained FCN-ACGAN network model, training a pre-constructed initial ResNet-LSTM classification model according to the real sample and the simulation sample to obtain an optimal classification model, classifying the terahertz time-domain spectrum data by using the optimal classification model, and determining the attribute and the type of the detected article according to a classification result.

Description

A method and device for identifying hidden dangerous goods based on FCN-ACGAN data enhancement

technical field

The present invention relates to the technical field of item identification, in particular to the identification of hidden dangerous goods by analyzing terahertz time-domain spectra, and in particular to a method and equipment for identifying hidden dangerous goods based on FCN-ACGAN data enhancement.

Background technique

Terahertz waves are widely used in the field of material detection because of their fingerprint spectrum. Using terahertz waves to identify various dangerous goods can achieve non-destructive testing. With the rapid development of artificial intelligence, deep learning is used to analyze terahertz time It is becoming more and more common to use the trained model for non-destructive testing. However, the accuracy of the model generated by deep learning is extremely dependent on the amount of training terahertz time-domain spectral data. Insufficient training data will reduce the performance of the model. Therefore, it is necessary to enhance the terahertz time-domain spectral data.

Some experts and scholars have conducted research on the enhancement of terahertz time-domain spectral data, mainly through the effective generative confrontation network (GAN) in the time-series data enhancement method of deep learning to expand the original spectral data, based on WGAN (Wasserstein GAN ) new generation of artificial intelligence small sample data enhancement method, first divide the original sample into training set and test set samples, use training set samples to train GAN and generate simulated sample data, and expand the size of training set samples; then, use simulated samples to train Get the classifier; finally, use the test set samples to test the classification effect of the classifier.

However, the use of generative confrontation network (GAN) to enhance terahertz time-domain spectral data also has certain limitations. The traditional GAN model is difficult to model time-series data, and there are problems of unstable training, gradient disappearance, excessive freedom of mode and easy collapse. , leading to the lack of characteristic information and poor representativeness of the final generated data. Furthermore, the classification effect of the classifier obtained by training the data with missing characteristic information is difficult to meet the detection requirements, resulting in low recognition accuracy for item recognition based on the classification results.

Contents of the invention

The present invention provides a hidden dangerous article identification method and equipment based on FCN-ACGAN data enhancement, which is used to solve the problem of poor classification effect of the classification model obtained by training due to insufficient training data in the existing data classification method, which further leads to poor article identification accuracy Low technical issues.

The present invention provides a terahertz time-domain spectrum hidden dangerous goods identification method based on FCN-ACGAN data enhancement, the method comprising:

Preprocess the pre-acquired real spectral data to obtain real samples;

Utilize the pre-trained FCN-ACGAN network model to generate simulated samples; wherein, the FCN-ACGAN network model includes a generator and a discriminator, and the generator and the discriminator respectively include several fully connected layers; train the FCN - The steps of the ACGAN network model are as follows:

Step S1, using the real samples to pre-train the discriminator to obtain a primary discriminator;

Step S2, using the generator to generate primary simulation samples;

Step S3, mixing the primary simulation samples with the real samples to obtain training samples;

Step S4, using the training samples to actually train the primary discriminator and the generator, update the network parameters of the primary discriminator and the generator based on the RMSProp optimizer, and judge the Describe whether the FCN-ACGAN network model reaches Nash equilibrium, if so, then stop training, obtain the trained FCN-ACGAN network model, if not, then return to step S2;

Training the pre-built initial ResNet-LSTM classification model according to the real sample and the simulated sample to obtain an optimal classification model;

The optimal classification model is used to classify the terahertz time-domain spectral data, and the hidden dangerous goods are determined according to the classification result.

Preferably, the pre-built initial ResNet-LSTM classification model is trained according to the real sample and the simulated sample, and the optimal classification model obtained is specifically:

mixing the real sample with the simulated sample to obtain an extended sample, and dividing the extended sample into a pre-training sample and an actual training sample;

Construct an initial ResNet-LSTM classification model, use the pre-training samples to pre-train the initial ResNet-LSTM classification model, obtain the hyperparameters of the initial ResNet-LSTM classification model, based on the hyperparameters and the actual training The sample performs actual training on the initial ResNet-LSTM classification model, and when the model error of the initial ResNet-LSTM classification model meets the preset error threshold, the actual training ends and the optimal classification model is obtained.

Preferably, the use of the generator to generate primary simulation samples is specifically:

A mapping relationship between the real data and random noise in the latent space is established, and the generator generates simulation samples based on the mapping relationship, wherein the simulation samples include category labels.

Preferably, the RMSProp-based optimizer updates the network parameters of the primary discriminator and the network parameters of the generator respectively as follows:

Keeping the network parameters of the generator unchanged, obtaining the network loss value of the primary discriminator, updating the primary discriminator based on the RMSProp optimizer and the network loss value of the primary discriminator, when the primary discriminator When the number of updates of the discriminator satisfies the preset first update threshold, the network parameters of the primary discriminator are kept unchanged, the network loss value of the generator is obtained, and the network loss value of the generator is based on the RMSProp optimizer and the generator. The generator is updated until the number of updates of the generator meets a preset second update threshold.

Preferably, the generator includes: an input module, a Dense fully connected layer, a Tanh activation function layer and an output module.

Preferably, the discriminator includes: an input module, a Dense fully connected layer, a ReLU activation function layer and a softmax classifier.

Preferably, before step S2, the method further includes: using the real samples to pre-train the initial generator to obtain a trained generator.

Preferably, the preprocessing of the pre-acquired real spectral data to obtain the real sample is specifically:

Perform data cleaning on the pre-acquired real spectral data to obtain the first initial sample;

performing missing value supplementation on the first sample to obtain a second initial sample;

Perform normalization processing on the second initial samples to obtain real samples.

Preferably, the classification results include a first classification result and a second classification result, and the determination of hidden dangerous goods according to the classification results is specifically:

Judging the type of the classification result, when the classification result is the first classification result, it is determined that the detected item is a non-concealed dangerous product, and when the classification result is the second classification result, it is determined that the detected item is hidden dangerous goods, and comparing the second classification result with a pre-established spectral feature database of hidden dangerous goods, and judging the type of hidden dangerous goods according to the comparison results.

The present invention also provides an electronic device, which is characterized in that it includes a memory and a processor, and a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the aforementioned data The steps of the classification method.

As can be seen from the above technical solutions, the present invention has the following advantages:

The present invention provides a method and device for identifying hidden dangerous goods based on FCN-ACGAN data enhancement. Based on the RMSProp optimizer and the pre-acquired real samples of the detected items, the primary discriminator and generator including several fully connected layers are processed. After training and updating, the FCN-ACGAN model that achieves Nash equilibrium is obtained. Among them, the fully connected layer can help the FCN-ACGAN model learn the dynamic characteristics between real data and improve the quality of the data generated by the generator. Using the RMSProp optimizer to update the network parameters of the discriminator and generator can effectively eliminate the update process. The jitter caused by the gradient difference speeds up the convergence process of the FCN-ACGAN model. Further, use the generator in the trained FCN-ACGAN network model to create simulated samples, mix simulated samples and real samples to train the initial ResNet-LSTM classification model, and obtain the optimal classification model that meets the data classification standards. The model realizes the classification of the terahertz time-domain spectral data of the detected items, and finally identifies the attributes and types of the detected items according to the classification results. The data classification method provided by the present invention adopts the FCN-ACGAN model to create simulation samples, realizes the expansion of training samples, and solves the problem that in the existing data classification methods, the classification effect of the classification model obtained by training is poor due to insufficient training data, which further leads to the accuracy of item recognition Low technical issues.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. For those skilled in the art, other drawings can also be obtained according to these drawings without any creative effort.

Fig. 1 is a method flowchart of a method for identifying hidden dangerous goods based on FCN-ACGAN data enhancement provided by an embodiment of the present invention;

Fig. 2 is a network structure diagram of the FCN-ACGAN model provided by the embodiment of the present invention;

FIG. 3 is a network structure diagram of a generator provided by an embodiment of the present invention;

FIG. 4 is a network structure diagram of a discriminator provided by an embodiment of the present invention.

Detailed ways

The embodiment of the present invention provides a method and device for identifying hidden dangerous goods based on FCN-ACGAN data enhancement. By training and updating the primary discriminator and generator including several fully connected layers, an FCN-ACGAN that achieves Nash equilibrium is obtained. Network model, use the generator in the FCN-ACGAN network model to create simulation samples, realize the expansion of training samples, and further use the extended samples to train the initial ResNet-LSTM classification model to obtain the optimal classification model that meets the data classification standards, which solves the existing problems In the data classification method, due to insufficient training data, the classification effect of the trained classification model is poor, which further leads to the technical problem of low accuracy of item recognition.

In order to make the purpose, features and advantages of the present invention more obvious and understandable, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the following The described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

The deep learning network model usually extracts the general characteristics of the data set, uses the characteristics as the characteristics of predicting a certain type of results, and uses the data set to train the constructed model, so that the output results of the model are as close as possible to the above characteristics. When the training data is small, the model obtained by this training method often has limitations: when the training set data is used on the model to predict the result, the prediction result is better, and the test set or verification set data is used on the model. When predicting the results, the error rate of the model is high and the generalization ability is low. In order to avoid the above problems, deep learning algorithms often need to obtain a large amount of training data, and avoid over-fitting of the model by increasing the amount of training data.

Terahertz time-domain spectroscopy technology uses terahertz pulses to reflect or transmit on the surface of the sample, respectively measure the reference signal and measurement signal before and after passing through the sample, and then transform the collected time-domain signal through fast Fourier transform (FFT) In the frequency domain, two frequency domain spectra are obtained. Finally, by processing the frequency domain data, the optical parameters such as the refractive index, extinction coefficient and absorption coefficient of the tested sample can be extracted. By analyzing the relevant data, the object detection can be realized ( item category). Existing technologies usually use terahertz time-domain spectral data as training samples to establish a deep learning network model, and then use the model to detect objects. The acquisition of domain spectral data is usually manually marked, which often requires a lot of labor and time costs, and the amount of data obtained cannot cover the actual object detection scene. Insufficient training data will lead to low recognition accuracy and universality of the deep learning classification model. As a result, the recognition rate of item attributes based on classification results is low.

In view of this, this application provides a method for identifying hidden dangerous goods based on FCN-ACGAN data enhancement, please refer to Figure 1, the method includes:

100. Perform preprocessing on the pre-acquired real spectral data to obtain real samples.

It is understandable that in the process of spectral data collection, there will inevitably be noise. Preprocessing the original terahertz time-domain spectral data of the detected item (hereinafter collectively referred to as data, will not be described in detail) can improve the expression of data. Ability to make data features more obvious.

200. Generate a simulation sample by using a pre-trained FCN-ACGAN network model.

The constructed model is trained with sample data to make the accuracy of the model as high as possible, which is the ideal result of model training. However, when the training data is small, the model obtained by this training method often appears that when the model uses the training set data to predict the result, the prediction result is better, and the test set or verification set data is used on the model to predict the result. , the error rate of the model is high. In order to avoid the above problems, the establishment of the model often needs to obtain a large amount of training data, and the phenomenon of over-fitting of the model can be avoided by increasing the amount of training data. When the amount of original data is small, the amount of data can be expanded by creating samples.

It is understandable that when there is a sufficient amount of training data, a model with high accuracy can be trained, but if the quality of the training data itself is poor and the training data is insufficient, the trained model usually has poor accuracy or cannot To meet the requirements of use, therefore, it is necessary to ensure that the training data is sufficient, and at the same time ensure that the sufficient training data has good quality.

When the amount of data is small, the total number of samples can be increased through data enhancement, but if the total number of samples is obtained first through data enhancement, and then the total data is processed to improve the data quality, the processing workload will increase. Therefore, this implementation The method firstly preprocesses the original data, and then based on the original data, uses the pre-trained FCN-ACGAN network model to generate simulation samples to achieve number sample expansion.

300. Train the pre-built initial ResNet-LSTM classification model according to the real samples and the generated samples to obtain an optimal classification model.

According to step 200, the number of simulated samples that meet the training needs can be obtained, the simulated samples are mixed with the original samples to realize the expansion of the training samples, and the ResNet-LSTM classification model is trained using the expanded samples. When the model error meets the preset error threshold, the training is stopped. Get the best classification model.

400. Use the optimal classification model to classify the terahertz time-domain spectral data, and determine hidden dangerous goods according to the classification result.

Preferably, in this embodiment, the classification results are divided into two types: the first classification result and the second classification result, wherein the first classification result indicates that the detected item is not dangerous, and the second classification result indicates that the detected item Dangerous goods hidden in the item

The details of determining hidden dangerous goods based on classification results are as follows:

First judge the type of the classification result. When the classification result is the first classification result, it is determined that the detected item is a non-concealed dangerous product. When the classification result is the second classification result, it is determined that the detected item is Concealing the dangerous goods, and comparing the second classification result with a pre-established spectral feature database of the hidden dangerous goods, and judging the type of the hidden dangerous goods according to the comparison result.

It is understandable that the terahertz spectral data of different items are different. By comparing the terahertz spectral features of the items with the established spectral feature databases of various dangerous goods, it is possible to judge whether the detected items are dangerous goods and what kind of goods they are. dangerous goods.

The data classification method provided by the present invention utilizes the pre-trained FCN-ACGAN network model to realize number sample expansion, and further uses the expanded samples to train the initial ResNet-LSTM classification model to obtain the optimal classification model that meets the data classification standard, and then utilizes the most The optimal classification model classifies the terahertz time-domain spectral data of the detected items, and determines the hidden dangerous goods according to the classification results, which solves the problem of poor classification effect of the trained classification model due to insufficient training data in the existing data classification methods, which further leads to The technical problem of low recognition rate of item attributes based on classification results.

On the basis of the foregoing embodiments, the present application provides another preferred embodiment, and step 100 can be specifically implemented in the following manner:

performing data cleaning on the pre-acquired real spectral data to obtain a first initial sample, then supplementing the first sample with missing values to obtain a second initial sample, and finally performing normalization processing on the second initial sample, get real samples.

It is understandable that before supplementing the data with missing values, the data is cleaned first, which can remove noise data and irrelevant data in the data set, improve data quality, and avoid increasing the workload of processing useless data. Further, by cleaning Complementing the missing values of the missing data can realize data repair, reduce the risk of bias, improve sample representativeness, and further standardize and normalize the repaired data, so that the data features are relatively consistent, and the impact of outstanding features on model training can be eliminated.

On the basis of the aforementioned embodiments, the present application provides another preferred embodiment. In step 200, the FCN-ACGAN network model includes a generator and a discriminator, and the generator and the discriminator respectively include several fully connected layers. Among them, the FCN -The steps of building the ACGAN network model are as follows:

Step S1, using the real samples to pre-train the discriminator to obtain a primary discriminator;

Step S2, using the generator to create a primary generation sample;

Step S3, mixing the primary generated sample with the real sample to obtain a mixed sample;

Step S4, using the mixed samples to actually train the primary discriminator and the generator, update the network parameters of the primary discriminator and the generator based on the RMSProp optimizer, and judge that the FCN- Whether the ACGAN network model reaches the Nash equilibrium, if so, stop the training, and obtain the trained FCN-ACGAN network model, if not, return to step S2;

Wherein, before step S2, it also includes: using the real samples to pre-train the initial generator to obtain the trained generator.

Referring to Fig. 2, wherein, the initial FCN-ACGAN model includes a discriminator (network) and a generator (network), which is different from a traditional GAN network. In the FCN-ACGAN model of this embodiment, a full The connection layer, through the fully connected layer, can help the FCN-ACGAN model learn the dynamic characteristics between real data, improve the quality of the data generated by the generator, and enhance the ability of the discriminator to identify the authenticity of the data. Please refer to FIG. 3 and FIG. 4 , FIG. 3 is a network structure diagram of a generator provided in this embodiment, and FIG. 4 is a network structure diagram of a discriminator provided in this embodiment.

Among them, the generator consists of 1 input module, 5 Dense fully connected layers, 4 Tanh activation function layers, and 1 output module. The input to the generator is random noise and labeled data, and the output is a "simulated" sample.

The discriminator consists of 1 input module, 5 Dense fully connected layers, 3 ReLU activation function layers and 1 softmax classifier. The input of the discriminator contains "simulated" samples and real samples of the output of the generator, and the output is the discriminative result with the label type.

It is understandable that the initial generator that has not been learned (trained) does not know the "appearance" of the real data at the beginning, and directly uses the untrained initial generator to create data, and the distribution of the created data is consistent with the distribution of real data If the data generated by the initial generator is directly mixed with the real data, and then the mixed data is sent to the discriminator for model training, the training process of the model will be increased. Therefore, in order to speed up the training process, it is necessary to let the initial generator perform imitation learning first, so that the initial generator has a certain imitation ability, and then use the data created by the trained generator to train the discriminator.

Similarly, if the data generated by the generator is directly mixed with the real data, and then the mixed data is sent to the discriminator for model training, since the discriminator does not know the "appearance" of the real data at the beginning, the mixed data enters the discriminator Finally, the discriminator cannot make accurate judgments, and the ability of the discriminator to distinguish between true and false data can only be improved through multiple updates. The ability grows slowly, so directly mixing real and fake data to train the model will increase the training process of the model. In order to speed up the training process of the FCN-ACGAN model, before using mixed data for model training in this embodiment, the discriminator is pre-trained with real data first, so that the discriminator has the ability to distinguish between real and fake data at an early stage.

In the actual training stage, the generator uses the random noise in the latent space to establish a mapping relationship with the real data distribution, generates several simulated samples with labels, and then mixes several simulated samples with real samples to obtain training data, and inputs the training data into the In the discriminator after pre-training. The discriminator uses the training data for training, obtains the network loss value of the discriminator, and updates the network parameters of the discriminator according to the network loss value of the discriminator and the RMSProp optimizer. Every time it is updated, the discriminator will become more "smart". . When the number of updates of the discriminator meets the first update threshold, keep the parameters of the discriminator unchanged, obtain the network loss value of the generator, and update the network parameters of the generator according to the network loss value of the generator and the RMSProp optimizer, Similarly, the generator will become "smarter" every time it is updated, until the number of updates of the generator meets the preset second update threshold. The above process is repeated continuously. The generator continuously generates new simulated samples that are more similar to real samples, and uses the new simulated samples and original data to train the discriminator. The discriminator is constantly updated to improve the discrimination ability, and the generator is also constantly updated. The generator and the discriminator are updated alternately until the entire FCN-ACGAN model reaches the Nash equilibrium. Wherein, during the updating process, the learning rate of the generator and the discriminator is a preset value. In this embodiment, there is no specific limitation on the update frequency of the generator and the discriminator, and those skilled in the art can set it as needed.

When the FCN-ACGAN model reaches the Nash equilibrium, it means that the simulated samples generated by the generator in the FCN-ACGAN model can already "deceive" the discriminator. It can be understood that the simulated samples generated by the generator have met the requirements for training data , the expression ability of the simulated samples is close to that of the real data.

In the above embodiment, by adding a fully connected layer to the generator network and the discriminator network, it can help the FCN-ACGAN model learn the dynamic characteristics between real data, improve the quality of the data generated by the generator, and enhance the discriminator to identify the authenticity of the data. The ability enables the generator to generate simulated samples that are more similar to the real data. At the same time, in the update process of the generator network and the discriminator, the RMSProp optimizer is used to update the network parameters of the discriminator and the generator, which can effectively Eliminate the jitter caused by the gradient difference during the update process, and speed up the model convergence process.

On the basis of the foregoing embodiments, the present application provides another preferred embodiment, and step 300 can be specifically implemented in the following manner:

When enough simulated samples are obtained through the generator, the simulated samples are mixed with real samples to obtain extended samples, and the extended samples are used to train the ResNet-LSTM classification model. When the model error meets the preset threshold, the training is stopped and the optimal model is obtained. Classification model, where the mixed samples are divided into training samples and actual training samples.

Further, using mixed samples to train the ResNet-LSTM classification model specifically includes: constructing an initial ResNet-LSTM classification model, using pre-trained samples to pre-train the initial ResNet-LSTM classification model, and obtaining the initial ResNet-LSTM classification model. The hyperparameters of the model, based on the hyperparameters and the actual training samples, the initial ResNet-LSTM classification model is actually trained, and when the model error of the initial ResNet-LSTM classification model meets the error threshold, the actual training is ended, Get the best classification model.

Further, the optimal classification model can be used to classify the terahertz time-domain spectral data.

In order to verify the feasibility of the FCN-ACGAN data-enhanced terahertz time-domain spectrum hidden dangerous goods identification method, this embodiment provides an example of the verification accuracy of dangerous goods identification based on the aforementioned optimal classification model.

Select a number of preprocessed real samples, simulated samples created by FCN-ACGAN, and extended samples mixed with simulated samples created by FCN-ACGAN and real samples. Among them, the numbers of real samples, simulated samples and extended samples are the same. The original samples, simulated samples and extended samples were respectively input into the ResNet-LSTM classification model for classification and recognition. The classification results are shown in Table 1. It can be seen from Table 1 that the performance of the simulated samples generated by FCN-ACGAN and the real samples in the ResNet-LSTM classification model are basically the same, which verifies that the simulated samples generated by the FCN-ACGAN model provided by this application can not only capture the effective features of the original data, It is also possible to generate new samples that fit the characteristics of the real data. The classification model has a recognition rate of 99.42% for the extended samples obtained by mixing simulated samples and real samples created by FCN-ACGAN, and a recognition rate of 98.33% for real samples. The recognition accuracy rate is increased by 1.09%, and the recognition accuracy of item recognition based on the classification results is improved, which verifies the feasibility and effectiveness of the terahertz time-domain spectrum hidden dangerous goods recognition method based on FCN-ACGAN data enhancement.

Table 1 Combined FCN-ACGAN and ResNet-LSTM recognition algorithm performance table

The present invention is based on the FCN-ACGAN data enhanced terahertz time-domain spectrum hidden dangerous goods identification method. After expanding the data set based on the FCN-ACGAN model, it can effectively improve the over-fitting problem caused by insufficient data and improve the classification model training. The final identification accuracy rate, so that the accuracy rate of hidden dangerous goods identification is higher.

The present application also provides an electronic device, the device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the aforementioned data classification method .

Those skilled in the art can clearly understand that for the convenience and brevity of the description, the specific working process of the above-described system, device and unit can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here.

In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods may be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or May be integrated into another system, or some features may be ignored, or not implemented. In another point, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be in electrical, mechanical or other forms.

The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, each unit may exist separately physically, or two or more units may be integrated into one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

If the integrated unit is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the essence of the technical solution of the present invention or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , including several instructions to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, and other media that can store program codes.

As mentioned above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it can still understand the foregoing The technical solutions recorded in each embodiment are modified, or some of the technical features are replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims (10)

1. A terahertz time-domain spectrum concealed hazardous article identification method based on FCN-ACGAN data enhancement is characterized by comprising the following steps:

preprocessing pre-acquired real terahertz spectrum data to obtain a real sample;

generating a simulation sample by using a pre-trained FCN-ACGAN network model; the FCN-ACGAN network model comprises a generator and a discriminator, wherein the generator and the discriminator respectively comprise a plurality of full connection layers; the steps of training the FCN-ACGAN network model are as follows:

s1, pre-training the discriminator by using the real sample to obtain a primary discriminator;

s2, generating a primary simulation sample by using the generator;

s3, mixing the primary simulation sample with the real sample to obtain a training sample;

s4, actually training the primary arbiter and the generator by using the training samples, respectively updating the network parameters of the primary arbiter and the network parameters of the generator based on a RMSProp optimizer, judging whether the FCN-ACGAN network model reaches Nash equilibrium, if so, stopping training to obtain a trained FCN-ACGAN network model, and if not, returning to the S2;

training a pre-constructed initial ResNet-LSTM classification model according to the real sample and the simulation sample to obtain an optimal classification model;

and classifying the terahertz time-domain spectral data by using the optimal classification model, and determining hidden dangerous goods according to a classification result.

2. The method for identifying the terahertz time-domain spectrum concealed hazardous articles based on FCN-ACGAN data enhancement as claimed in claim 1, wherein the training of the pre-constructed initial ResNet-LSTM classification model according to the real samples and the simulation samples to obtain the optimal classification model specifically comprises:

mixing the real sample and the simulation sample to obtain an extended sample, and dividing the extended sample into a pre-training sample and an actual training sample;

constructing an initial ResNet-LSTM classification model, utilizing the pre-training sample to pre-train the initial ResNet-LSTM classification model to obtain a super parameter of the initial ResNet-LSTM classification model, actually training the initial ResNet-LSTM classification model based on the super parameter and the actual training sample, and ending the actual training when the model error of the initial ResNet-LSTM classification model meets a preset error threshold to obtain an optimal classification model.

3. The method for identifying the terahertz time-domain spectroscopy concealed hazardous articles based on FCN-ACGAN data enhancement as claimed in claim 2, wherein the generating of the primary simulation sample by the generator specifically comprises:

establishing a mapping relationship between the real data and random noise of a potential space, and generating a simulation sample based on the mapping relationship, wherein the simulation sample contains a class label.

4. The method for identifying the terahertz time-domain spectroscopy concealed hazardous article based on FCN-ACGAN data enhancement as claimed in claim 3, wherein the updating of the network parameters of the primary discriminator and the generator respectively based on the RMSProp optimizer comprises:

keeping the network parameters of the generator unchanged, acquiring a network loss value of the primary discriminator, updating the primary discriminator based on the RMSProp optimizer and the network loss value of the primary discriminator, keeping the network parameters of the primary discriminator unchanged when the updating times of the primary discriminator meet a preset first updating threshold, acquiring the network loss value of the generator, and updating the generator based on the RMSProp optimizer and the network loss value of the generator until the updating times of the generator meet a preset second updating threshold.

5. The FCN-ACGAN data enhancement based terahertz time-domain spectroscopy concealed hazardous article identification method according to claim 4, wherein the generator comprises: the device comprises an input module, a Dense full-connection layer, a Tanh activation function layer and an output module.

6. The FCN-ACGAN data enhancement based terahertz time-domain spectroscopy concealed hazardous article identification method according to claim 5, wherein the discriminator comprises: the device comprises an input module, a Dense full connection layer, a ReLU activation function layer and a softmax classifier.

7. The method for identifying the concealed hazardous article based on the FCN-ACGAN data enhancement terahertz time-domain spectroscopy as claimed in claim 6, further comprising, before step S2: and pre-training the initial generator by using the real sample to obtain a trained generator.

8. The method for identifying the terahertz time-domain spectrum concealed hazardous article based on FCN-ACGAN data enhancement as claimed in claim 1, wherein the preprocessing is performed on the pre-acquired real spectrum data to obtain the real sample specifically as follows:

performing data cleaning on pre-acquired real spectrum data to obtain a first initial sample;

supplementing missing values to the first sample to obtain a second initial sample;

and carrying out normalization processing on the second initial sample to obtain a real sample.

9. The method for identifying the concealed hazardous article based on the FCN-ACGAN data enhancement terahertz time-domain spectroscopy as claimed in claim 1, wherein the classification result comprises a first classification result and a second classification result, and the determining the concealed hazardous article according to the classification result specifically comprises:

and judging the type of the classification result, judging that the detected object is a non-hidden dangerous article when the classification result is a first classification result, judging that the detected object is a hidden dangerous article when the classification result is a second classification result, comparing the second classification result with a pre-established spectral feature database of the hidden dangerous article, and judging the type of the hidden dangerous article according to the comparison result.

10. An electronic device, comprising a memory and a processor, wherein the memory has stored thereon a computer program, which, when executed by the processor, causes the processor to carry out the steps of the data classification method according to any one of claims 1 to 9.

Images