Tire trace identification method based on progressive mask and multistage feature fusion

Abstract

The invention discloses a tire trace identification method based on progressive mask and multistage feature fusion, and belongs to the technical field of image identification, and the method comprises the steps: 1, obtaining a tire trace stained image, and carrying out the preprocessing of the image, and obtaining a preprocessed image; 2, inputting the preprocessed image into an encoder layer to obtain encoding features; 3, inputting the coding features into a decoder layer to obtain decoding features; 4, post-processing the decoded features to obtain a reconstructed trace image; 5, inputting the reconstructed trace image into an encoder layer in the retrieval model to obtain semantic features of the trace; and 6, classifying the tire trace images according to the semantic features of the traces. According to the method, the reconstruction model is utilized to reconstruct the stained pattern trace of the tire, the stained area is repaired, and the category label of the reconstructed image can be accurately retrieved, so that the efficient retrieval of the tire trace image is realized, and the intelligent development of tire trace recognition is promoted.

Description

Technical Field

[0001] The present invention belongs to computer vision technology in the field of artificial intelligence, and specifically relates to a tire mark image recognition method based on progressive mask and multi-level feature fusion, which can be widely used in tread analysis in the tire industry and identification and tracking of tire marks in the field of criminal investigation. Background Art

[0002] Tire pattern is one of the most important features of tires. Its structural characteristics not only affect the mechanical properties of vehicles such as grip, noise, rolling resistance, etc., but also play a huge role in resolving commercial copyright disputes. There are many types of tire patterns in the market, and tire patterns have the characteristics of rich texture, clear edges, and similar visual effects of tire patterns of different models. These factors have posed new challenges to the high reliability of intelligent detection of tire patterns. Therefore, how to effectively classify tire patterns and make reasonable pattern similarity judgments is of great significance for solving practical problems. At the same time, cars have become the main vehicles for criminals to commit crimes or roam. In recent years, criminals' anti-reconnaissance awareness has been continuously enhanced. License plates, as the most direct feature for tracking vehicles, are no longer reliable. It is urgent to establish a comprehensive description of the characteristics of the vehicle involved in the accident. Unlike other features of the vehicle, tire marks, as indirect features of the vehicle, are left at the scene as fixed evidence and are often ignored by criminals. They have important identification value. Therefore, it is of great significance for today's science and technology-based police work to quickly check the vehicles at the crime scene and track the vehicles that caused the accident, study the tire mark images, and improve the intelligence level of trace identification.

[0003] Tire tread trace retrieval is a special branch in the field of computer vision. The relevant retrieval algorithms can be divided into traditional algorithms based on artificial feature extraction and neural network algorithms based on data drive. Traditional algorithms rely on artificial feature extractors, are sensitive to data environments, and are difficult to adapt to complex situations. They generally have problems such as difficulty in data feature extraction, complex mathematical operation steps, poor generalization, and poor recognition accuracy. Although algorithms based on neural networks can achieve higher retrieval accuracy, they are limited by the lack of data sets in the tire field and generally have the problem of overfitting sample data. At the same time, in complex environments, the acquired tread trace features are usually incomplete or unclear, and neural networks are black box models. It is difficult for humans to observe their feature extraction process, especially when it comes to cases such as violent crimes and commercial copyright disputes. Inappropriate retrieval results will cause controversy and public opinion. Therefore, the main problem at present is that the process of making judgments on on-site images is not clear enough, resulting in the current subjective experience of experts as the judgment standard. The feature extraction algorithm adopted does not have adaptive capabilities, has a low degree of intelligence, and the retrieval efficiency is difficult to meet application requirements. Summary of the invention

[0004] To address the deficiencies and drawbacks in the above-mentioned existing technologies, the present invention proposes a tire mark recognition method based on progressive masking and multi-level feature fusion, which can break through the dataset limitations, achieve intelligent reconstruction and recognition of tire tread marks, and quickly and efficiently complete the retrieval of common tire marks.

[0005] To achieve the above objective, the technical solution adopted by the present invention is as follows:

[0006] A tire mark recognition method based on progressive masking and multi-level feature fusion according to the present invention is characterized by including the following steps:

[0007] Step 1: After obtaining the tire mark image and performing data augmentation, a tire mark image dataset is obtained. , denotes the th enhanced tire mark image, denotes the number of tire mark images, denotes the number of channels of the tire mark image, denotes the height of the tire mark image, denotes the width of the tire mark image; let The class label of ;

[0008] Step 2: Construct a reconstruction network including: a block embedding layer , a block encoder layer , a channel masking layer , a channel encoder layer , a latent space layer , a skip connection layer , a block decoder layer , a channel decoder layer , a linear layer , a block feature mapping layer and a channel feature mapping layer ;

[0009] Step 3: Use the block embedding layer , the channel masking layer , the block encoder and the channel encoder to construct the block encoded feature of the th tire mark image , the channel encoded feature ;

[0010] Step 4: Use the linear layer , the latent space layer and the skip connection layer , the block decoder and the channel decoder , the block feature mapping layer and the channel feature mapping layer pair and are processed to obtain the detailed tire mark image and the reconstructed tire mark image ;

[0011] Step 5: Calculate the total loss according to the reconstruction result , which is used for backpropagation, and the network weights are updated using the optimizer until convergence, so as to obtain the trained reconstruction model ;

[0012] Step 6: Based on the trained construct a retrieval network and fine-tune it to obtain the retrieval model , which is used for classifying and identifying the input tire mark image.

[0013] Another feature of the tire mark image reconstruction and recognition method based on progressive masking and multi-level feature fusion according to the present invention is that the , and , , in step 2 are all stacked by the same modules, and each module is composed of a multi-head self-attention module and a fully connected feed-forward network, and a residual operation layer and a regularization operation layer are connected after each module;

[0014] The block embedding layer includes: a two-dimensional convolutional layer and a normalization layer, where the convolutional kernel size of the two-dimensional convolutional layer is , being the convolution stride;

[0015] The channel mask layer includes: a squeeze-and-excitation module, a normalization layer, an activation function, and a Gumbel-Softmax layer;

[0016] The linear layer , the skip connection layer , the block feature mapping layer and the channel feature mapping layer are all single linear layers.

[0017] Furthermore, step 3 includes the following steps:

[0018] Step 3.1: The Zhang tire track image Input block embedding layer is processed to obtain the th patched feature . Then, after unfolding and transposing the last two dimensions of , the th trace embedding feature is obtained , where represents the number of patches in and represents the dimension of a single patch in

[0019] Step 3.2: The input channel mask layer is processed to obtain the channel mask matrix of ;

[0020] Step 3.3: Add two-dimensional sine-cosine positional encoding to to obtain the th trace feature with positional information ;

[0021] Step 3.4: Perform random block masking on to obtain the th masked trace embedding feature , the position recovery index of the patches in and the block mask matrix of , where represents the number of remaining patches after random block masking, and , r represents the masking ratio, and ;

[0022] Step 3.5: Perform a forward operation on input block encoder to obtain the block encoding feature of ;

[0023] Step 3.6: After performing a Hadamard product operation on and , perform a forward operation on input channel encoder to obtain the channel encoding feature of .

[0024] Furthermore, the random block masking in Step 3.4 includes the following steps:

[0025] Step 3.4.1: Based on 's shape, sample a group of random noises that follow a uniform distribution , and obtain 's order index sorted from smallest to largest, denoted as 's patch disorder index ;

[0026] Step 3.4.2: Obtain 's order index sorted from largest to smallest, denoted as 's patch position recovery index ;

[0027] Step 3.4.3: Based on the mask ratio , intercept the first elements in to obtain 's patch mask index , where K represents the number of elements, and ;

[0028] Step 3.4.4: Using as the index, extract the corresponding elements in to obtain the rd masked trace embedding feature ;

[0029] Step 3.4.5: Based on 's shape, create a mask matrix with all element values being zero , set the first element values in to 1, and replicate along the last dimension times to obtain a new mask matrix ;

[0030] Step 3.4.6: Using as the index, reorder the elements in to obtain 's block mask matrix .

[0031] Furthermore, the said Step 4 includes the following steps:

[0032] Step 4.1: Input into the linear layer for forward operation to obtain the th dimension-reduced trace encoding feature , where represents 's dimension of a single patch;

[0033] Step 4.2: Construct a first masked token with all elements being zero and a second masked token ;

[0034] Step 4.3: According to the elements in as indices, restore the patch order after concatenating and in the row direction, so as to obtain the latent space input feature ; ;

[0035] Step 4.4: Input into the latent space layer and perform a forward operation to obtain the latent space reconstruction feature of ; ;

[0036] Step 4.5: Input into the skip connection layer and perform a forward operation to obtain the th dimension-reduced block encoding feature of ;

[0037] Step 4.6: According to the elements in as indices, restore the patch order after concatenating and in the row direction, so as to obtain the conditional input feature of ;

[0038] Step 4.7: Sum and element-wise and input into the block decoder to perform a forward operation to obtain the block decoding feature of ;

[0039] Step 4.8: Input into the channel decoder and perform a forward operation to obtain the channel decoding feature of : :

[0040] Step 4.9: Use Equation (1) and Equation (2) to obtain the tire tread detail image and the tire tread reconstruction image :

[0041] (1)

[0042] (2)

[0043] In formulas (1) and (2), represents the forward operation of represents the forward operation of represents the inverse transform operation of the discrete cosine transform, represents the inverse patching operation for reshaping the tensor shape.

[0044] Furthermore, step 5 includes the following steps:

[0045] Step 5.1: Using as the reconstruction target, construct a reconstruction loss function using formula (3) :

[0046] (3)

[0047] In formula (3), represents element-wise multiplication, represents the row index of the patched feature; represents the patching operation that only changes the tensor shape;

[0048] Step 5.2: Construct a detail loss function for the tire track image using formula (4) :

[0049] (4)

[0050] In formula (4), represents Gaussian blur, represents the sharpening operation;

[0051] Step 5.3: Construct the total loss using formula (5) :

[0052] (5)

[0053] In formula (5), represents the weight coefficient of is the weight coefficient of

[0054] Furthermore, step 6 includes the following steps:

[0055] Step 6.1: Remove the channel mask layer, latent space layer, skip connection layer, block decoder layer, channel decoder layer, linear layer, block feature mapping layer, and channel feature mapping layer in and add a linear classification layer to construct a retrieval network ;

[0056] Step 6.2: Using input for forward processing in to obtain the predicted class features , where represents the number of classes;

[0057] Step 6.3: Calculate the cross-entropy loss according to and for backpropagation of the retrieval network , and use optimizer to update the weights of the retrieval network until the cross-entropy loss converges, so as to obtain the trained retrieval model .

[0058] An electronic device according to the present invention includes a memory and a processor, characterized in that the memory is used to store a program for supporting the processor to execute the tire mark recognition method, and the processor is configured to execute the program stored in the memory.

[0059] A computer-readable storage medium according to the present invention, characterized in that the computer program stored on the computer-readable storage medium executes the steps of the tire mark recognition method when run by a processor.

[0060] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0061] 1. The present invention designs an algorithm based on a masked autoencoder architecture, adopting a progressive masking strategy and multi-level feature fusion, which can realize the reconstruction of tire mark damaged images and the imaging of detailed features of tire marks, thus significantly reducing the difficulty of tire mark identification and improving the intelligent level of tire mark identification.

[0062] 2. The present invention adopts an unsupervised learning method, relying on the reconstruction task as a pretext task, without relying on artificial class labels, breaking through the dataset limitation of tire mark images, and being able to quickly migrate to related downstream tasks with the help of the general representation of tire mark images, which can not only significantly improve the applicability of the algorithm in actual scenarios, but also achieve high-precision retrieval of tire mark images.

[0063] 3. The present invention adopts the method of "first reconstructing the tire mark damaged image and then retrieving the tire mark image category", demonstrating the decision-making process of the algorithm, which is helpful for practitioners to quickly verify the retrieval results. Especially in the criminal investigation field, the present invention can significantly reduce the public opinion problems caused by the opaque decision-making of the retrieval method. Description of the Drawings

[0064] Figure 1This is the standard process for the structure and training of the reconstruction model in the present invention;

[0065] Figure 2 This is the schematic diagram of the minimum unit structure in the present invention;

[0066] Figure 3 This is the reconstruction example diagram of some tire trace damaged images in the example of the present invention;

[0067] Figure 4 This is the schematic diagram of the inference process of the present invention. Specific embodiments

[0068] The present invention will be further described below in conjunction with the accompanying drawings and specific examples.

[0069] In this example, as Figure 1 shown, in a tire trace image recognition method based on progressive masking and multi-level feature fusion, the training process of the reconstruction model includes the following steps:

[0070] Step 1: Obtain tire trace images and perform data augmentation to obtain an instance dataset , represents the th image, represents the number of images, where represents the number of channels of the image, represents the height of the image, represents the width of the image; in this example, taking the tire trace image dataset as an example, the image instance shape is unified as , and the instance dataset has a total of 5600 image data, that is, ;

[0071] Set the class labels of the tire trace image instance dataset to construct a label instance dataset , where, represents the th image instance corresponding class label. In this example, there are a total of 70 types of tire trace images, with 80 tire trace images for each type. Among them, .

[0072] Step 2: Construct a reconstruction network including: a block embedding layer , a block encoder layer , a channel masking layer , a channel encoder layer , a latent space layer , a skip connection layer , a block decoder layer , Channel decoder layer , Linear layer , Block feature mapping layer and channel feature mapping layer .

[0073] Among them, the block embedding layer contains a two-dimensional convolutional layer and a normalization layer. The convolutional kernel size of the convolutional layer is , and the two-dimensional discrete cosine transform matrix is used as the initialization matrix. The stride of the convolution is . In this example ; , and , , are all stacked by the same modules. Each module consists of a multi-head self-attention module and a fully connected feed-forward network, and a residual operation layer and a regularization operation layer are connected after each module;

[0074] Channel mask layer contains: squeeze-and-excitation module, normalization layer, activation function and Gumbel-Softmax layer;

[0075] Linear layer , Skip connection layer , Block feature mapping layer and channel feature mapping layer are all single linear layers.

[0076] This example is based on the Python language and the Pytorch architecture, uses ViT as the backbone model, and constructs the network by adopting the method of modular programming. The schematic diagram of the minimum unit structure is as shown in Figure 2 . The hyperparameter settings of each part of the structure in this example are shown in Table 1;

[0077] Table 1 Schematic diagram of the encoder, decoder, latent space layer and their minimum unit structures and hyperparameter settings in this example

[0078]

[0079] Step 3, construct the block encoding feature instance and channel encoding feature instance of the th tire track image instance ;

[0080] Step 3.1, input the th tire track image into the block embedding layer for patching processing to obtain the patched characteristics , expand to obtain the th trace embedding feature instance , where represents the number of patches in after block embedding, represents the dimension of a single patch in , ;

[0081] Step 3.2, input into the input channel mask layer for processing to obtain the channel mask matrix instance .

[0082] Step 3.3, add two-dimensional sine-cosine positional encoding to according to Equations (1), (2), and (3) to obtain the th trace feature instance with positional information ;

[0083] (1)

[0084] (2)

[0085] (3)

[0086] In the equations, represents the index of the patch. In this instance, is , representing the positional index of the patch block in the row or column direction, where , are respectively the row and column sizes of the feature after patching, and , ; represents the index in the direction; in this instance , .

[0087] Step 3.4, perform random block masking on to obtain the th masked trace embedding feature instance , the instance of the positional recovery index of the patches in and the block mask matrix instance , where Indicates the number of patches remaining after the random block mask, satisfying , Indicates the mask ratio. In this example , .

[0088] Step 3.4.1: Based on the shape, sample a set of random noise that follows a uniform distribution , and obtain the order index of the elements in sorted from smallest to largest, denoted as the disorder index instance of the patches in

[0089] Step 3.4.2: Obtain the order index of the elements in sorted from largest to smallest, denoted as the position recovery index instance of the patches in

[0090] Step 3.4.3: Based on the mask ratio , intercept the elements in , retain the first elements in .

[0091] Step 3.4.4: Using as the index, extract the elements at the corresponding positions in to obtain the th masked trace embedding feature instance

[0092] Step 3.4.5: Based on the shape, create a mask matrix instance with all element values being zero , set the first element values to 1, and using as the index, reorder the elements in , and replicate 768 times along the last dimension in to obtain the block mask matrix instance .

[0093] Step 3.5: Input into the block encoder for a forward operation to obtain the block encoding feature instance :

[0094] Step 3.6: Combine with After performing the Hadamard product operation, the input channel encoder performs a forward operation to obtain the channel-encoded feature instance .

[0095] Step 4. Decode and to obtain the tire track detail image instance and the tire track reconstruction image instance ;

[0096] Step 4.1. Input into the linear layer to perform a forward operation to obtain the th dimensionality-reduced trace-encoded feature instance ;

[0097] Step 4.2. Construct the first masked token instance and the second masked token instance .

[0098] Step 4.3. Concatenate and in the row direction, and then restore the patch order in the concatenated feature using the elements in as indices to obtain the latent space input feature instance ;

[0099] Step 4.4. Input into the latent space layer to perform a forward operation to obtain the latent space reconstruction feature instance .

[0100] Step 4.5. Input into the skip connection layer to perform a forward operation to obtain the th dimensionality-reduced block-encoded feature instance ;

[0101] Step 4.6. Concatenate and in the row direction, and then restore the patch order in the concatenated feature using the elements in as indices to obtain the input conditional feature instance ;

[0102] Step 4.7. Input the element-wise sum of and into the block decoder Perform a forward operation in to obtain the block decoding feature instance of ;

[0103] Step 4.8, input into the channel decoder and perform a forward operation to obtain the channel decoding feature instance of .

[0104] Step 4.9, use equations (4) and (5) to obtain the detail image instance of the tire track and the reconstructed image instance of the tire track :

[0105] (4)

[0106] (5)

[0107] In equations (4) and (5), represents the forward operation of , represents the forward operation of , represents the inverse discrete cosine transform operation, represents the inverse patchification operation for reshaping the feature shape, specifically reshaping the feature tensor originally in the shape of into an image in the shape of ; Part of the reconstructed image of this instance is shown in Figure 3 as follows. From left to right, they are: the original tire track image, the tire track mask image, the tire track reconstructed image, and the tire track detail image.

[0108] Step 5, calculate the total loss for backpropagation of , and update the network weights using optimizer until converges, thus obtaining the trained reconstruction model ;

[0109] Step 5.1, use as the reconstruction target and construct the reconstruction loss function using equation (6):

[0110] (6)

[0111] In equation (6), represents the element-wise product operation, represents the row index of the feature after patchification;​ A patching operation that does not change the internal elements of the feature, that is, reshaping an image with a shape of into a feature tensor with a shape of .

[0112] Step 5.2, constructing a detail loss function for the tire track image using Equation (7) :

[0113] (7)

[0114] In Equation (7), represents Gaussian blur, represents a sharpening operation.

[0115] Step 5.3, constructing the total loss using Equation (8) :

[0116] (8)

[0117] In Equation (8), represents 's weight coefficient, is 's weight coefficient;

[0118] In this example, the training of the retrieval network in a tire track image recognition method based on progressive masks and multi-level feature fusion includes the following steps:

[0119] Step 6, constructing a retrieval network based on the trained to obtain a retrieval model for classifying and recognizing the input tire track image;

[0120] Step 6.1, constructing a retrieval network based on the structure, including: a block embedding layer , a block encoder layer , a channel encoder layer , a linear classification layer , where the linear classification layer instance includes an average pooling layer and a linear layer instance, and the network layers in

[0121] Step 6.2, performing forward processing with as the input in to obtain the predicted class feature instance of ;

[0122] Step 6.3, according to and Calculate the cross - entropy loss for backpropagation in the retrieval network and use the optimizer to update the weights of the retrieval network until the cross - entropy loss converges, thus obtaining the trained retrieval model .

[0123] The complete retrieval process of this example is as Figure 4 shown, and the inference steps include:

[0124] Step 7: For a single damaged tire - mark image Obtain the mask image of the tire mark through manual masking ;

[0125] Step 8: Input into to obtain the reconstructed image of the tire mark , and conduct manual evaluation. If the main features in are complete, it is used as the subsequent input; otherwise, needs to be re - masked and input into for reconstruction;

[0126] Step 9: Input into to obtain the predicted class features ;

[0127] Step 10: Take the maximum value in as the predicted class result to obtain the class prediction result of ; ;

[0128] Step 11: Manually verify whether the tire - mark image corresponding to is consistent with . If it is consistent, then the confidence ; if it is inconsistent, then query and compare with in turn according to the top five results in the value of ; if it is still inconsistent, the manual identification result needs to be taken as the final result.

[0129] In this embodiment, an electronic device includes a memory and a processor. The memory is used to store a program that supports the processor to execute the above - mentioned method, and the processor is configured to execute the program stored in the memory.

[0130] In this embodiment, a computer - readable storage medium stores a computer program, and when the computer program is run by a processor, it executes the steps of the above - mentioned method.

Claims

1. A tire mark recognition method based on progressive masking and multi-level feature fusion, characterized in that: The steps include: Step 1: Obtain tire trace images and perform data enhancement to obtain a tire trace image dataset , Indicates the enhanced images of tire tracks, represents the number of tire track images, represents the number of channels of the tire track image, represents the height of the tire track image, represents the width of the tire track image; let The category label is recorded as ; Step 2: Build the reconstruction network Includes: Block Embedding Layer , block encoder layer , channel mask layer , channel encoder layer , hidden space layer , skip connection layer , block decoder layer , channel decoder layer , Linear layer , block feature map layer With channel feature map layer ; Step 3: Using Block Embedding Layer , channel mask layer , block encoder and channel encoder Build Tire track images Block encoding features , channel coding features ; Step 4: Using the Linear Layer , hidden space layer and skip connection layers , Block Decoder and channel decoder , block feature map layer With channel feature map layer right and Process and obtain Detail image of tire tracks and tire track reconstruction images ; Step 5: Calculate the total loss based on the reconstruction results , used for Perform back propagation and use The optimizer updates the network weights until Converge to obtain the reconstructed model after training ; Step 6: Based on the trained Building a retrieval network And fine-tune to get the retrieval model , used to classify and identify the input tire track image.

2. The tire track image reconstruction and recognition method based on progressive mask and multi-level feature fusion according to claim 1, characterized in that: In step 2 , and , , They are all stacked by the same modules. Each module consists of a multi-head self-attention module and a fully connected feed-forward network, and each module is connected to a residual operation layer and a regularization operation layer. Block Embedding Layer Contains: a two-dimensional convolutional layer and a normalization layer, where the convolution kernel size of the two-dimensional convolutional layer is , is the stride of the convolution; Channel Mask Layer Contains: compression excitation module, normalization layer, activation function and Gumbel-Softmax layer; Linear Layer , skip connection layer , block feature map layer With channel feature map layer Each is a single linear layer.

3. The tire track image reconstruction and recognition method based on progressive mask and multi-level feature fusion according to claim 2, characterized in that: Step 3 includes the following steps: Step 3.1: Tire track images Input Block Embedding Layer is processed and the Patched features , and then After expanding and transposing the last two dimensions of Traces Embedding Features ,in, express The number of patches in express The dimensions of a single patch in ; Step 3.2: Input channel mask layer Processed in The channel mask matrix ; Step 3.3: Add two-dimensional sine-cosine position coding to obtain the first Trace features ; Step 3.4: Perform random block masking to obtain The masked trace embedding features , The location of the patch is restored index and The block mask matrix ,in, represents the number of patches remaining after random block masking, and , r represents the mask ratio, and ; Step 3.5: Input Block Encoder Perform the forward operation in Block encoding features ; Step 3.6: and After the Hadamard product operation, the input channel encoder Perform the forward operation in Channel coding characteristics .

4. The tire track image reconstruction and recognition method based on progressive mask and multi-level feature fusion according to claim 3, characterized in that: The random block mask in step 3.4 includes the following steps: Step 3.4.1, based on The shape of the sample is a set of random noises that obey a uniform distribution. , and obtain The sequential index of the elements in the array is sorted from small to large, denoted by Disordered index of patches in ; Step 3.4.2, obtain The order index of the elements in the sequence from large to small is recorded as The location of the patch is restored index ; Step 3.4.3: Based on mask ratio , intercept Center front elements, and get Mask index of the patch in , where K represents the number of elements, and ; Step 3.4.4: For index, extract The elements at the corresponding positions in Masked trace embedding features ; Step 3.4.5: Based on The shape of the mask matrix is ​​created with all element values ​​​​set to zero ,Will Center front Set the element value to 1 and Copy along the last dimension times, and obtain a new mask matrix ; Step 3.4.6: For index, Rearrange the elements in to get The block mask matrix .

5. The tire track image reconstruction and recognition method based on progressive mask and multi-level feature fusion according to claim 4, characterized in that: The step 4 comprises the following steps: Step 4.1: Input Linear Layer Perform forward operation in The trace encoding features after dimensionality reduction ,in, express The dimensions of a single patch in ; Step 4.2: Construct the first masked word whose elements are all zero and the second mask word ; Step 4.3: According to The element in is the index, and The patch order after splicing in the row direction is restored to obtain The latent space input features ; Step 4.4: Input latent space layer Perform the forward operation in Latent space reconstruction features ; Step 4.5: Input skip connection layer Perform forward operation in Block encoding features after dimensionality reduction ; Step 4.6: The element in is the index, and The patch order after splicing in the row direction is restored to obtain Conditional input features ; Step 4.7: and Element-wise summation is then fed into the block decoder Perform the forward operation in Block decoding features ; Step 4.8: Input channel decoder Perform the forward operation in Channel decoding characteristics : Step 4.9: Use equation (1) and equation (2) to obtain tire trace detail images and tire track reconstruction images : (1) (2) In formula (1) and formula (2), express The forward operation, express The forward operation, represents the inverse transform operation of discrete cosine transform, Represents an inverse patching operation that reshapes a tensor.

6. The tire track image reconstruction and recognition method based on progressive mask and multi-level feature fusion according to claim 5, characterized in that: The step 5 comprises the following steps: Step 5.1: As the reconstruction target, the reconstruction loss function is constructed using formula (3): : (3) In formula (3), represents the element-wise product operation, Indicates the row index of the patched feature; represents a patching operation that only changes the shape of a tensor; Step 5.2: Use formula (4) to construct the detail loss function of the tire track image : (4) In formula (4), represents Gaussian blur, Indicates a sharpening operation; Step 5.3: Use formula (5) to construct the total loss : (5) In formula (5), express The weight coefficient of for The weight coefficient of .

7. The tire track image reconstruction and recognition method based on progressive mask and multi-level feature fusion according to claim 6, characterized in that: The step 6 comprises the following steps: Step 6.1, removal The channel mask layer, latent space layer, skip connection layer, block decoder layer, channel decoder layer, linear layer, block feature map layer and channel feature map layer in , and add a linear classification layer , thereby constructing a retrieval network ; Step 6.2: enter Perform forward processing in The predicted category features ,in, Indicates the number of categories; Step 6.3, according to and Calculate the cross entropy loss for the retrieval network Perform back propagation and use Optimizer Update Retrieval Network The weights of are calculated until the cross entropy loss converges, thus obtaining the trained retrieval model. .

8. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the tire mark recognition method according to any one of claims 1 to 7, and the processor is configured to execute the program stored in the memory.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the tire mark recognition method according to any one of claims 1 to 7 are executed.

Images