A method and system for identifying the authenticity of a product based on a key identification site

Abstract

The application provides a kind of based on key identification site's commodity true or false identification method and system, obtain the key identification site image of commodity and obtain data set, the image in data set is artificially identified true or false classification, and obtain true or false classification data set;Resnet50 is constructed based on key identification site's true or false classification network, including main network, the regional peak elimination strategy inserted in main network and the feature interaction module connected at the output end of main network;Loss function is constructed, and the true or false classification network based on key identification site is trained using true or false classification data set, determines the parameter of true or false classification network, and obtains the true or false classification network model based on key identification site;Shielding regional peak elimination strategy in the true or false classification network model based on key identification site, and the image of goods to be detected is input into the true or false classification network model based on key identification site and carries out true or false identification classification, and the application true or false identification accuracy is high, and practicality is strong.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of commodity authenticity identification, and specifically belongs to a commodity authenticity identification method and system based on key identification parts. BACKGROUND

[0002] Traditional commodity authenticity identification methods require certain professional knowledge, rich experience and specific instruments. At present, expert identification method is still the mainstream commodity authenticity identification method. For example, the authenticity of a stamp is identified by ink marks and text printing displacement, and the authenticity of calligraphy is identified by local subtle differences and seal material. However, the workload of comparing all areas of the work with the authentic product one by one is very large, which easily leads to a decrease in the accuracy of the identification result.

[0003] Thanks to the rapid development of deep learning and computer vision, authenticity identification based on image recognition technology has become a research trend. For example, a microscopic identification method of traditional Chinese medicinal materials and decoction pieces is adopted by using a microscope to extract samples and using a convolutional neural network to extract important structural feature characteristics. A characteristic analysis and determination is completed by using deep learning after a sample of a traditional Chinese medicinal material specimen is directly taken under a low-power microscope by using a CCD camera. For example, a crocodile skin accessory authenticity identification method is adopted by extracting the grain surface tissue structure and skin shape of the crocodile skin accessory. For example, a computer vision and machine learning-based authenticity cigarette packaging identification model is obtained by simulating human visual perception through computer vision, solving the problems of low efficiency and strong subjectivity in manual authenticity identification of cigarettes, and improving the identification efficiency and adaptability of authentic and fake cigarettes. For example, a deep learning image classification technology is used to identify and classify the authenticity of corn varieties, solving the problems of subjective judgment, time-consuming and laborious, and poor practicability in traditional mode. For example, a painting authenticity auxiliary identification method is adopted by using a twin network structure to compare the feature similarity between a representative work of a painter and a painting to be identified. For example, a calligraphy authenticity identification method is adopted by using an invariant moment in the pattern recognition field to evaluate the shape authenticity, which can be used as an auxiliary basis for identifying the authenticity of calligraphy works, and improve the efficiency and credibility of calligraphy identification.

[0004] However, the existing authenticity identification methods based on image recognition technology have great limitations. (1) Most of the designed networks only target certain specific commodities, and their generalization and authenticity identification performance in other commodity fields are not high. (2) The focus of authenticity identification is to find discriminative features, and the existing methods have poor ability to locate discriminative features. The features focused by the network are often not the most discriminative features. SUMMARY

[0005] In order to solve the problems in the prior art, the application provides a commodity authenticity identification method and system based on key identification parts, acquires key identification part images, constructs a corresponding data set, adopts Resnet50 as a backbone network, proposes a regional peak elimination strategy, designs a feature interaction module on the basis, and constructs a commodity authenticity identification network based on key identification parts, relies on a visual classification algorithm to extract features of key identification parts that can best distinguish authenticity, and performs authenticity identification, which can be widely used for authenticity identification of various commodities and has high authenticity identification accuracy.

[0006] To achieve the above object, the application provides the following technical scheme: a commodity authenticity identification method based on key identification parts, and the specific steps are as follows:

[0007] S1 acquires key identification part images of commodities to obtain a data set, artificially identifies and classifies authenticity of images in the data set to obtain an authenticity classification data set;

[0008] S2 constructs an authenticity classification network based on key identification parts, including Resnet50 as a backbone network, a regional peak elimination strategy RPE inserted in the backbone network, and a feature interaction module FIM connected at an output end of the backbone network;

[0009] S3 constructs a loss function, trains the authenticity classification network based on key identification parts by using the authenticity classification data set, determines parameters of the authenticity classification network, and obtains an authenticity classification network model based on key identification parts;

[0010] S4 shields the regional peak elimination strategy in the authenticity classification network model based on key identification parts, and inputs an image of a commodity to be detected into the authenticity classification network model based on key identification parts to perform authenticity identification and classification.

[0011] Further, in S2, a regional peak elimination strategy RPE is inserted between Resnet50 backbone networks stage3 and stage4 and between stage4 and stage5, and outputs of stage3 and stage4 are taken as inputs of the regional peak elimination strategy RPE.

[0012] Further, in S2, the regional peak elimination strategy RPE has the following specific operation steps:

[0013] 1) an input feature map F is obtained, and the input feature map F is expressed as F element of R C×W×H Slices are performed on the input feature map F along a width dimension to obtain n sliced feature maps F (k) ∈R C ×(W / n)×H ,k element of [1, n], each sliced feature map is subjected to channel pooling processing to obtain a feature map

[0014] 2) select feature map F P(k) The pixel with the maximum peak value in F is set as a threshold δ, and an elimination mask P is generated according to the size of the pixel and the threshold δ (k)drop ;

[0015] 3) multiply the elimination mask P (k)drop with the feature map F to obtain F (k)out , and splice back to the CxWxH dimension according to the width dimension to obtain the peak elimination feature F result ;

[0016] Further, in S2, if the pixel in the feature map F P(k) is greater than the threshold δ, the elimination mask P (k)drop is set to 0, and conversely the elimination mask P (k)drop is set to 1.

[0017] Further, in S2, the two feature maps obtained by the backbone network after Stage4 and Stage5 are taken as an image pair (F1, F2), and the image pair (F1, F2) is the input of the feature interaction module FIM.

[0018] Further, in S2, the specific operation steps of the feature interaction module FIM are as follows:

[0019] 1) compress the feature maps F1, F2 in width and height from W1xH1, W2xH2 to L1, L2 to obtain

[0020] 2) perform inner product operation on the feature maps F′1 T and F2′ to obtain a similarity matrix M 12 , and perform normalization operation on the row and column of the interaction matrix of the similarity matrix M 12 to obtain an interaction feature map W 12 ;

[0021] 3) weight the interaction feature map W 12 to the feature maps F′1, F2′ according to formulas (7) and (8), and convert the sizes L1, L2 back to W1xH1, W2xH2 to obtain a correlation feature map

[0022] W F1 = F2′xW 12 T + F1′ (7)

[0023] W F2 = F1′xW 12 + F2′ (8)

[0024] 4) again splice the correlation feature map W F1down-sampling and correlation feature map W F2 Element-wise addition finally outputs the addition result.

[0025] Further, in S2, the feature interaction module FIM is connected to the full connection layer to perform classification operation, to obtain the probability of the goods being true and false, and the sum of the scores of the two is 1, when the probability of the goods being true is greater than the probability of the goods being false, the image to be detected is considered to be classified as true.

[0026] The application provides a key identification part-based goods true and false identification system, comprising:

[0027] A data acquisition module is configured to acquire key identification part images of goods to obtain a data set, perform artificial identification and true and false classification on the data set, and obtain a true and false classification data set.

[0028] A network construction module is configured to construct a true and false classification network based on key identification parts, comprising a Resnet50 backbone network, a regional peak elimination strategy RPE inserted in the backbone network, and a feature interaction module FIM connected to an output end of the backbone network.

[0029] A network training module is configured to construct a loss function, train the true and false classification network based on key identification parts by using the true and false classification data set, determine parameters of the true and false classification network, and obtain a true and false classification network model based on key identification parts.

[0030] The regional peak elimination strategy in the true and false classification network model based on key identification parts is shielded, and a to-be-detected goods image is input into the true and false classification network model based on key identification parts for true and false identification classification.

[0031] The application provides a terminal device, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps in the true and false identification method when executing the computer program, or the processor implements the functions of the modules in the system when executing the computer program.

[0032] The application further provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps in the true and false identification method.

[0033] Compared with the prior art, the application has at least the following beneficial effects:

[0034] The application provides a commodity authenticity identification method based on key identification parts, and the authenticity classification network based on the key identification parts takes Resnet50 as a backbone network, a regional peak elimination strategy RPE is proposed in the network training stage, the most significant information is inhibited, potential distinguishable information is forced to be captured, a feature interaction module FIM is proposed on this basis, commonality of various features is established, feature richness is improved, and therefore the most distinguishable features on the key identification parts are obtained to identify authenticity, authenticity identification accuracy is improved, and the method can be widely used in various commodities and has high practicability.

[0035] In existing concealed and non-concealed anti-counterfeiting identification technologies, holograms, bar codes, special anti-counterfeiting labels and other technologies are applied to products when the products are manufactured, and the application is a practical and non-invasive identification method, does not need to embed any substance in the products, and does not need to modify the product objects in any way; and the existing concealed and non-concealed anti-counterfeiting identification technologies cannot verify previously unlabeled objects, which also means that the above method needs to label each product, if the label is damaged or the like, the object cannot be identified, and the method can distinguish authenticity according to subtle differences in image identification parts, without the need to label the products.

[0036] The method can complete identification work in the cloud when used, a user only needs to upload the photographed image according to a prompt, without the need for any additional equipment, and the cost is more low, and the process is simple and convenient. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 Structure diagram of the authenticity classification network based on the key identification parts;

[0038] Figure 2 Structure diagram of the regional peak elimination strategy RPE;

[0039] Figure 3 Structure diagram of the feature interaction module FIM;

[0040] Figure 4 Authenticity images of the inner stamping parts of a Louis Vuitton high-end handbag (a) (b) (c) authentic product; (d) (e) (f) fake product;

[0041] Figure 5 Authenticity images of the metal zipper parts of a Louis Vuitton high-end handbag (a) (b) (c) authentic product; (d) (e) (f) fake product;

[0042] Figure 6 Feature map visualization effect diagram. DETAILED DESCRIPTION

[0043] The application will be further described in combination with the drawings and specific embodiments.

[0044] As Figure 1 illustrated, the present application provides a method for identifying the authenticity of goods based on key identification sites, and the specific steps are as follows:

[0045] S1. Constructing a true and false classification data set;

[0046] Obtain the data set by acquiring the image of the key identification site of the goods, and divide the data set into a test set and a training set. The images in the training set are artificially identified by professional appraisers to classify true and false, and true and false data sets are obtained;

[0047] The key identification site is the site on the goods that can best distinguish true and false or the site in the judgment basis set. Different goods can determine different key identification sites according to their characteristics. For example, the key identification sites of Louis Vuitton high-end handbags are the inner standard stamp site and the metal zipper site.

[0048] S2. Constructing a true and false classification network based on key identification sites. Resnet50 is used as the backbone network. According to the identification process of small local features in human eyes, a regional peak elimination strategy (RPE) is proposed in the network training stage. The overall feature map is cut into small slices, and the true and false information is explored in each slice to simulate the true and false exploration process of the local features of the identification object by the appraiser in the artificial identification process, suppress the most prominent information, force the network to capture potential distinguishable information, and obtain peak elimination features F result with potential true and false characteristics. After obtaining the peak elimination features F result , a feature interaction module (FIM) is used to enhance the semantic complementary information, improve the relevance, establish the common relationship of each feature, improve the feature richness, and enhance the clue comparison ability.

[0049] S2.1 Backbone network

[0050] As Figure 1 illustrated, the present application selects Resnet50 as the backbone network. This network can be described as 5 stages. The stage1 structure is simple and can be regarded as the preprocessing of the input image. The last 4 stages are composed of residual blocks (Bottleneck), and the structures are similar. The network depth increases with the increase of the stage, and the contained information also becomes more and more rich. When the input image passes through different stages, feature maps F∈R C×W×H of different scales can be obtained, where C, W and H represent the channel number, width and height of the feature map.

[0051] S2.2 Regional peak elimination strategy

[0052] This invention aims to prevent the true / false classification network from focusing solely on the most salient appearance features, while guiding the network to consider subtle features that distinguish between true and false. To comprehensively capture these subtle features, this invention inserts a Region Peak Elimination (RPE) strategy between stages 3, 4, and 5 of the ResNet50 backbone network. The outputs of stages 3 and 4 are used as inputs to the RPE strategy. The RPE slices the feature map, dividing it into multiple segments, and explores subtle features within each segment, thus avoiding overemphasis on the most salient appearance features. Figure 2 As shown, the specific structure of the Region Peak Elimination Strategy (RPE) is as follows:

[0053] First, for the input feature map F∈R C×W×H Slicing along the width dimension yields n slice feature maps F. (k) ∈R C×(W / n)×H k∈[1,n], where the computational complexity of the slicing operation is proportional to the number of slices. If n is too large, the computational load will be too large and the applicability will be reduced. Conversely, if n is too small, the exploration of fine regions will be insufficient. Therefore, based on actual engineering experience, this invention sets it to 7.

[0054] Then, each slice feature map is processed by channel pooling according to formula (1) to obtain... For classification tasks, The intensity of each pixel represents its ability to distinguish.

[0055]

[0056] To avoid focusing on the most salient features and ignoring other discriminative details, this invention eliminates the most discriminative parts, according to F. P(k) The pixel with the largest peak value is set as the threshold δ, and the elimination mask P is generated based on the size of the pixel and the threshold δ. (k)drop As shown in formula (2), for the elimination mask, the pixels greater than the threshold are set to 0, and the pixels less than the threshold are set to 1.

[0057]

[0058] Eliminate mask P (k)drop Feature maps obtained from slicing operations Multiply to get F (k)out And by concatenating the data back to the C×W×H dimension along the width dimension, the peak elimination feature F is obtained. result Specifically, as shown in formulas (3) and (4).

[0059]

[0060] F result = concat(F (k)out ) (4)

[0061] GAP in the above formula (1) represents global average pooling, and concat in formula (4) represents splicing of each feature map after being processed in the width dimension. It should be noted that in order to ensure the stability of the verification result in the inference stage, the strategy is only applied in the training process, and the strategy is not used after the network training is completed.

[0062] S2.3 Feature Interaction Module

[0063] After the above regional peak elimination strategy is completed, the identifiable peak elimination features F result captured need to be associated and the clue comparison ability needs to be enhanced. Therefore, the feature interaction module FIM is designed, which will model the interaction of peak elimination features F result at different levels, force different layers of the real-fake classification network to share the information mined, and enhance the comparison ability of each real-fake feature clue, promote the network to comprehensively judge the real-fake attribute. As shown in the following formula (5), the specific structure of the feature interaction module FIM is as follows: Figure 3

[0064] Firstly, the feature maps after the backbone network passes through Stage4 and Stage5 are taken as an image pair (F1, F2), wherein the channel number of the feature map F1 after convolution layer processing is the same as that of the feature map F2. In order to avoid destroying the original spatial information, the width and height dimensions of the feature map F1 remain unchanged, that is, the width and height of the two feature maps are compressed from W1xH1, W2xH2 to L1, L2 to obtain Then, the inner product operation is performed on the feature map F′1 T and the feature map F2′ to obtain a similarity matrix M 12 , wherein an element M i,j in the similarity matrix represents the similarity between the i th th pixel of the feature map F′1 and the j th th pixel of the feature map F2′. Each pixel contains the spatial features of the target object. When the similarity between two pixels is lower, the information they pay attention to is more different, and the complementarity is stronger. Therefore, -M 12 is taken as an interaction matrix, and a row and column normalization operation is performed on the interaction matrix to obtain an interaction feature map W 12 , which is specifically shown in the following formulas (5) and (6).

[0065] M 12 = F′1 T ​F2' (5)

[0066]

[0067] Secondly, the interaction feature map W 12 According to formula (7) (8) weighted to F'1, F2', the next step is to convert the size L1, L2 back to W1xH1, W2xH2, and get the correlation feature map Correlation feature map W F1 , W F2 The basic features of feature maps F1, F2 and the interaction information of F1 and F2 are contained in the correlation feature map.

[0068] W F1 =F2'xW 12 T +F1' (7)

[0069] W F2 =F1'xW 12 +F2' (8)

[0070] Finally, the correlation feature map W F1 is down-sampled and added to the correlation feature map W F2 element by element, and the sum result is finally output, and a full connection layer is used to complete the classification operation, and the probabilities of true and false of the goods are obtained through the classifier, and the scores of the two are added to 1, when the probability of true is greater than the probability of false, the image to be measured is considered to be classified as true.

[0071] S3 constructs a loss function, trains the true and false classification network to obtain a true and false classification network model based on the key identification part;

[0072] The loss between the sample label y and the predicted probability distribution is calculated by using the loss function, which can ensure that the true and false classification network classifies the images to be measured as true as much as possible, and classifies the images to be measured as false, so as to obtain a trained true and false classification network model based on the key identification part.

[0073] The present application uses cross entropy as the loss function of the above-mentioned true and false classification network, and iteratively optimizes the parameters in the true and false classification network by minimizing the loss function, and the loss function is as follows:

[0074]

[0075] Wherein is the probability that the sample of the true and false classification network is true, and y is the sample label, if the sample is true, the value is 1, otherwise the value is 0.

[0076] S4 carries out the commodity authenticity identification

[0077] The trained key identification part-based authenticity classification network model is used for commodity authenticity identification, and in the identification, the region peak elimination strategy is shielded, the key identification part image of the test set of S1 or the commodity to be detected is input into the key identification part-based authenticity classification network model, the probability that the commodity belongs to true and false can be obtained respectively, when the probability that the commodity is true is greater than the probability that the commodity is false, the image to be detected is considered to be true, otherwise, it is false.

[0078] The application also provides a key identification part-based commodity authenticity identification system, comprising:

[0079] A data acquisition module is configured to acquire the key identification part image of the commodity to obtain a data set, artificially identify the authenticity classification of the data set, and obtain an authenticity classification data set;

[0080] A network construction module is configured to construct a key identification part-based authenticity classification network, including a Resnet50 backbone network, a region peak elimination strategy RPE inserted in the backbone network, and a feature interaction module FIM connected to the output end of the backbone network;

[0081] A network training module is configured to construct a loss function, train the key identification part-based authenticity classification network by using the authenticity classification data set, determine the parameters of the authenticity classification network, and obtain a key identification part-based authenticity classification network model;

[0082] The region peak elimination strategy in the key identification part-based authenticity classification network model is shielded, and the image of the commodity to be detected is input into the key identification part-based authenticity classification network model for authenticity identification and classification.

[0083] The application provides a terminal device, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps in the above method are implemented. Alternatively, when the processor executes the computer program, the functions of the modules / units in the above system are implemented.

[0084] The above computer program can be divided into one or more modules / units, which are stored in the above memory and executed by the processor to complete the application.

[0085] The terminal device can be a desktop computer, a notebook computer, a palm computer, a cloud server, and other computing devices.

[0086] The terminal device can include, but is not limited to, a processor and a memory.

[0087] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component, and the like.

[0088] The memory can be used to store the above computer programs and / or modules, and the processor realizes the various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory, and calling the data stored in the memory.

[0089] The modules / units integrated in the terminal device, if realized in the form of software function units and sold or used as independent products, can be stored in a computer readable storage medium.

[0090] Based on such understanding, the present application realizes all or part of the above method, which can also be completed by instructing related hardware through a computer program. The computer program can be stored in a computer readable storage medium, and the computer program can realize the above method steps when executed by a processor. The computer program includes computer program codes, which can be in the form of source code, object code, executable files or some intermediate forms, etc.

[0091] The above computer readable medium can include any entity or device, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. that can carry the above computer program codes. It should be noted that the content included in the above computer readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, the computer readable medium does not include electric carrier signals and telecommunication signals.

[0092] Embodiments

[0093] Experimental environment setting

[0094] The laboratory environment of the present application adopts Ubuntu 16.04 system, the GPU model is RTX2080Ti, and the experimental environment configuration is CUDA 10.2+anaconda3+python3.6+pytorch-1.12.0. In order to fully train the model, the size of the input image is adjusted to 550*550 during the training process, and is randomly cropped to 448*448, and the sample image is randomly horizontally flipped, so as to enhance the robustness of the CNN model. The model uses cross entropy function to calculate the loss between the true label y and the predicted probability distribution, and updates the weight through back propagation, uses stochastic gradient descent method for optimization, the momentum is 0.9, the momentum attenuation is 0.00001, the epoch is set to 200, the batch is 20, the initial learning rate is set to 0.002, and the cosine annealing is used to adjust the learning rate. In the test process, the size of the input image is adjusted to 550*550, and the center is cropped to 448*448, and the accuracy is often used as an evaluation index for classification tasks.

[0095] Dataset

[0096] Taking a high-end LOUIS VUITTON handbag as an example, professional appraisers can identify its authenticity according to the gloss of the zipper head front and back, as well as the physical characteristics of the text on the inner standard stamp and the text on the external logo.

[0097] From Figure 4 It can be seen that the inner standard stamp of the LOUIS VUITTON bag will have the 'Louis Vuitton' word, and for the inner standard stamp part, the counterfeit often cannot be as same as the real product in terms of texture and depth, and identifying the authenticity from the text stamp of this part is one of the methods commonly used by appraisers and one of the more reliable identification methods.

[0098] Careful observation of the real and fake inner standard stamp part image shows that the real product stamp font is relatively shallow, and each letter is average and clear. This is because the real product is manually stamped by artificial pressure on each letter, while the counterfeit product will stamp the 'Louis Vuitton' word as a whole in order to save labor costs, which results in the letter texture of the counterfeit stamp being relatively deep, and the overall feeling of the stamp being obvious. Figure 5 It can be seen that the metal zipper of the real LOUIS VUITTON bag presents a bronze color, and the color is dark and dull, while the counterfeit presents a bright gold color, and the reflectivity is very strong. At the same time, the real product metal characters are uniform in thickness and uniform in depth, while the counterfeit metal characters are rough and uneven in depth;

[0099] The present application refers to the above-mentioned artificial identification process to design two key identification area positions of Louis Vuitton high-end handbags, i.e. the inner label stamping position and the zipper position, collect the images of the inner label stamping position and the zipper position of the Louis Vuitton high-end handbags, and then manually identify the images uploaded by professional appraisers, i.e. the professional appraisers make true and false classification for the collected images, and construct the data set as shown in Table 1:

[0100] Table 1 data set information

[0101]

[0102] For the inner label stamping, the present application will train and test according to the ratio of 8:2, i.e. a total of 7253 pictures, of which 80% is the training set, a total of 5802 pictures, and 20% is the test set, a total of 1451 pictures. At the same time, in order to avoid the problem of data imbalance, i.e. if there are a large number of counterfeits and only a small number of true products in the data set, the model will tend to the majority of counterfeit samples, which will eventually lead to the decline of classification performance. Therefore, in the 7253 inner label stamping position true and false images of the Louis Vuitton brand, 3626 are counterfeit product images and 3627 are true product images, so that the true and false quantity is close to 1:1 for training. For the zipper position data set, it is also divided in the same way as the inner label stamping, a total of 6048, of which 3024 are true products and 3024 are counterfeits.

[0103] 3.3 Classification network comparison experiment results

[0104] According to the above specific parameter settings, experiments are carried out, and the accuracy of each algorithm under different data sets is recorded in detail and the backbone network used is indicated.

[0105] The inner label stamping and zipper data set is divided into true and false two categories, the inner label stamping position and the zipper position are the identification positions commonly used by appraisers, through the characteristics of the lines, depth, luster, etc. of the position, the appraiser can judge the true and false of a handbag, which is one of the reliable identification means. The inner label stamping position and zipper position images collected by the present application are divided into true product images and counterfeit product images by professional appraisers.

[0106] Table 2 accuracy comparison of different algorithms on inner label stamping and zipper data set

[0107]

[0108] On the inner label embossed data set and the zipper data set, the application first carries out experiments on several commonly used backbone networks for classification tasks, as shown in Table 2, Vgg16, Resnet18, Resnet34, Resnet50, with the increase of network complexity, the performance is also improved, among these main stems, Resnet50 has stronger fitting ability for the true and false classification task, and excellent classification effect is obtained on the inner label embossed and zipper parts, and among many classification tasks, Resnet50 is also used as the leading backbone network, therefore, the application uses it as the backbone network to complete the feature extraction of the true and false images.

[0109] According to the uniqueness of the data set and the characteristics of the artificial identification process, the application designs the regional peak elimination strategy RPE and the feature interaction module FIM, and it can be found from Table 2 that on the inner label embossed data set, the application method finally achieves an accuracy of 95.5%, which is 4.3%, 3.3%, 2.7%, 2.7%, 2.2%, 1.5%, 1.4% higher than BCNN, MAMC, Mask-CNN, NTS, DTB-Net, API-Net, LIO under the condition of using the same backbone network Resnet50, which shows that the two modules designed by the application have excellent performance and better compatibility in the true and false classification network of the application.

[0110] On the zipper part data set, the performance of many algorithms decreases compared with the inner label embossed data set, because the zipper part data set has a more complex background, and the complex background information in the data set cannot provide valuable information for identification, but will increase the difficulty of accurate identification. As shown in Table 2, the application method achieves an accuracy of 93.1% on this data set, which is 5.8%, 5.2%, 5.1%, 3.4%, 2.5%, 0.7%, 2.0% higher than BCNN, MAMC, Mask-CNN, NTS, DTB-Net, API-Net, LIO.

[0111] According to the experimental results on the above two data sets, the effectiveness of the algorithm proposed by the application is fully verified, and the effectiveness of the true and false classification method based on the key identification part image of the application in the true and false identification task is illustrated.

[0112] Ablation experiment

[0113] Setting threshold δ

[0114] In RPE, the threshold δ in formula (2) determines the size of the peak region eliminated in the feature map. In order to discuss the influence of the threshold δ on the overall network performance, the present application constructs a key identification site-based authenticity classification network by combining RPE and FIM according to the above process with Resnet50 as the backbone, and carries out related experiments on the internal standard imprint and zipper dataset, and the results are recorded in Table 3.

[0115] In the regional peak elimination strategy, if the threshold δ is set too small, too many elimination regions will be caused, which will adversely affect the network performance. Therefore, the present application takes the threshold δ 0.5 as the initial value, and gradually increases the threshold δ to observe its influence on the classification result. It can be seen from Table 3 that on the internal standard imprint and zipper site dataset, when the threshold δ increases from 0.5 to 0.85, the accuracy also gradually increases, and after 0.85, the threshold δ continues to increase, and the accuracy begins to decrease, because when δ is too large, the eliminated region is too small and cannot achieve the expected effect. According to the results in Table 3, finally when the threshold δ is 0.85, the network performance reaches the best, and the accuracy is 95.5% on the internal standard imprint and 93.1% on the zipper site.

[0116] Table 3 Experimental results of threshold δ on internal standard imprint and zipper dataset

[0117]

[0118] Performance analysis of different modules

[0119] In order to respectively discuss the influence of the regional peak elimination strategy RPE and the feature interaction module FIM on the final classification performance, the present application carries out ablation experiments on the internal standard imprint and zipper dataset, and the results are recorded in Table 4.

[0120] Table 4 Experimental results of different modules on internal standard imprint and zipper dataset

[0121]

[0122] The accuracy of the original Resnet50 on the inner label printing dataset is 90.5%, and the accuracy on the zipper part dataset is 86.3%. According to the above description, and introducing the regional peak elimination strategy RPE based on Resnet50, the accuracy on the inner label printing and zipper datasets is 93.7% and 91.1% respectively, which is increased by 3.2% and 4.8% compared with the original baseline. In order to observe the effectiveness of the FIM module, remove the regional peak elimination strategy RPE and use the feature interaction module FIM in the backbone network alone, which is increased by 0.9% and 1.5% based on the baseline of 90.5% and 86.3%. When the regional peak elimination strategy RPE is combined with the feature interaction module FIM, the accuracy is 95.5% and 93.1%, which is increased by 5.0% and 6.8% compared with the baseline. It can be seen that the use of the feature interaction module FIM alone is lower than the use of the regional peak elimination strategy RPE alone, because the use of the feature interaction module FIM alone is not reasonable for the overall structure of the network. The reasonable way should be to use the feature interaction module FIM to establish the common relationship of each feature and improve its feature richness on the basis of capturing potential distinguishable features by the regional peak elimination strategy RPE, so as to improve the overall classification performance of the network. The results in Table 4 also further illustrate the effectiveness of the feature interaction module FIM and the regional peak elimination strategy RPE proposed in the present application on the inner label printing and zipper part datasets.

[0123] Visual analysis

[0124] In order to more intuitively understand the role of the two modules designed by the present application for the true and false classification task, and further enhance the credibility of the proposed method, the present application carries out a visual experiment on the inner label printing and zipper datasets, and uses visual analysis on the feature maps of the original Resnet50 and the method of the present application, and the results are shown in Figure 6

[0125] As can be seen from the figure, the backbone network Resnet50 model only focuses on a small number of appearance features and ignores some local distinguishable features, for example, in the zipper part dataset, only the metal zipper head is focused on, and the logo printing in the zipper is ignored, in the inner label printing dataset, only part of the significant seal font features are focused on, and more potential subtle local information is ignored. The method of the present application under the joint action of the regional peak elimination strategy and the feature interaction module can focus on more ignored true and false discrimination information in addition to focusing on significant features.

[0126] ​The method determines that the specific identification parts of the handbag are the internal standard imprint part and the zipper part, images of the two specific identification parts are collected, corresponding data sets are constructed, Resnet50 is used as a backbone network, a regional peak elimination strategy is proposed, and a feature interaction module is designed on this basis, and finally, the accuracy of 95.5% is achieved on the internal standard imprint, and the accuracy of 93.1% is achieved on the zipper part.

Claims

1. A method for authenticating a product based on a key authentication site, characterized by, The specific steps are as follows: S1, obtaining a data set by acquiring an image of a key identification part of a commodity, manually identifying and classifying the images in the data set to obtain a true and false classification data set; S2, constructing a true and false classification network based on the key identification part, including a Resnet50 backbone network, a regional peak elimination strategy RPE inserted in the backbone network, and a feature interaction module FIM connected at the output end of the backbone network; S3, constructing a loss function, training the true and false classification network based on the key identification part using the true and false classification data set, determining the parameters of the true and false classification network, and obtaining a true and false classification network model based on the key identification part; S4, shielding the regional peak elimination strategy in the true and false classification network model based on the key identification part, and inputting the image of the commodity to be detected into the true and false classification network model based on the key identification part for true and false identification classification; In S2, a regional peak elimination strategy RPE is inserted between stage3 and stage4 and between stage4 and stage5 of the Resnet50 backbone network, and the outputs of stage3 and stage4 are used as the inputs of the regional peak elimination strategy RPE; In S2, the specific operation steps of the regional peak elimination strategy RPE are as follows: 1) on the input feature map slicing operation in the width dimension obtains n sliced feature maps , channel pooling processing is performed on each sliced feature map to obtain a feature map , C, W, H represent the number of channels, width and height of the feature map 2) select features The pixel with the largest peak in the center is set as the threshold A removal mask is generated according to the size of the pixel and the threshold ;​ 3) eliminate mask with feature map multiplication , and concatenate back dimension to get peak elimination feature ; In S2, the specific operation steps of the feature interaction module FIM are as follows: 1) Combine the two feature maps obtained from Stage 4 and Stage 5 of the backbone network into a single image pair. The image pairs The feature map is the input to the Feature Interaction Module (FIM). In width and height by Compress to ,get ; 2) feature map with inner product operation to obtain a similarity matrix , the row and column of the interaction matrix of the similarity matrix are normalized to obtain an interaction feature map ; 3) the interaction feature map is weighted to the feature map according to formula (7) (8) is converted back to get the associated feature map ; (7) (8) 4) the associated feature map down-sampled and the associated feature map element-wise addition finally outputs the addition result.

2. The method according to claim 1, wherein S2, the feature map if the pixel is greater than the threshold then the mask is eliminated set to 0, otherwise the mask is eliminated set to 1.

3. The method according to claim 1, wherein In S2, the feature interaction module FIM is connected with a fully connected layer for classification operation, and the probabilities of the commodity being true and false are obtained, and the sum of the scores of the two is 1. When the probability of the commodity being true is greater than the probability of the commodity being false, the image to be detected is considered to be classified as true.

4. A product authenticity verification system based on a key identification site, characterized by, It comprises: The data acquisition module is used for acquiring an image of a key identification part of a commodity to obtain a data set, manually identifying and classifying the data set to obtain a true and false classification data set; The network construction module is used for constructing a true and false classification network based on the key identification part, including a Resnet50 backbone network, a regional peak elimination strategy RPE inserted in the backbone network, and a feature interaction module FIM connected at the output end of the backbone network; The network training module is used for constructing a loss function, training the true and false classification network based on the key identification part using the true and false classification data set, determining the parameters of the true and false classification network, and obtaining a true and false classification network model based on the key identification part; The network training module is used for constructing a loss function, training the true and false classification network based on the key identification part using the true and false classification data set, determining the parameters of the true and false classification network, and obtaining a true and false classification network model based on the key identification part; In the network construction module, a regional peak elimination strategy RPE is inserted between stage3 and stage4 and between stage4 and stage5 of the Resnet50 backbone network, and the outputs of stage3 and stage4 are used as the inputs of the regional peak elimination strategy RPE; In the network construction module, the specific operation steps of the regional peak elimination strategy RPE are as follows: 1) on the input feature map slicing operation in the width dimension to obtain n sliced feature maps channel pooling processing on each sliced feature map to obtain a feature map , C, W, H represent the number of channels, width and height of the feature map 2) select features The pixel with the largest peak in the center is set as the threshold A removal mask is generated according to the size of the pixel and the threshold ;​ 3) multiply with the mask elimination and feature map and concatenate back in width dimension dimension to get the peak elimination feature ;​ In the network construction module, the specific operation steps of the feature interaction module FIM are as follows: 1) Combine the two feature maps obtained from Stage 4 and Stage 5 of the backbone network into a single image pair. The image pairs The feature map is the input to the Feature Interaction Module (FIM). In width and height by Compress to ,get ; 2) feature map with inner product operation to obtain a similarity matrix , the row and column of the interaction matrix of the similarity matrix are normalized to obtain an interaction feature map ; 3) convert the interaction feature maps to feature maps according to equations (7) (8) with weights convert back to get the associated feature maps ; (7) (8) 4) the associated feature map down-sampled and the associated feature map element-wise addition finally outputs the addition result.

5. A terminal device, characterized by, It comprises: A processor, a memory, and a computer program stored in the memory and executable on the processor, the processor implementing the steps of the authentication method according to any one of claims 1-3 when executing the computer program, or the processor implementing the functions of the modules of the system according to claim 4 when executing the computer program.

6. A computer-readable storage medium, characterized in that, The computer program stored in the computer readable storage medium is executable on the processor, and the processor implements the steps of the authentication method according to any one of claims 1-3 when executing the computer program.

Images