An image target recognition method based on self-attention and low-rank decomposition
By employing self-attention and low-rank decomposition image target recognition methods, the problems of insufficient accuracy and robustness in weak target detection are solved, enabling effective detection and tracking in complex scenes and harsh environments. This method is applicable to anti-missile systems, maritime surveillance, and field reconnaissance.
Patent Information
- Application Number
- CN202310645084.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-01
- Publication Date
- 2025-11-25
- Estimated Expiration
- 2043-06-01
AI Technical Summary
Existing technologies lack accuracy and robustness in detecting small targets, especially in complex scenes and harsh environments where they are difficult to detect and track effectively. Traditional methods have reached a bottleneck, and deep learning methods are insufficient in improving detection performance.
An image target recognition method based on self-attention and low-rank decomposition is adopted. By generating a pseudo-label set, training a neural network and performing loss calculation, the self-attention module and low-rank decomposition technology are used for target recognition. The feasible region is determined by combining an interpretable self-attention module and the inverse IPI algorithm to ensure the reliability and interpretability of the recognition results.
It improves the accuracy and robustness of weak target detection, ensures the reliability and interpretability of identification results, and is applicable to scenarios such as anti-missile systems, sea surface monitoring, and field reconnaissance.
Smart Images
Figure CN116645588B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image target recognition technology, and more specifically to an image target recognition method based on self-attention and low-rank decomposition. Background Technology
[0002] Target detection and tracking remain important methods due to their low cost and resistance to interference. Especially in missile defense systems, maritime surveillance, and field reconnaissance, all potential targets should be detected as early as possible. Therefore, the detection of weak targets is a crucial step in target detection and tracking. Because weak targets lack texture and color information, possess only captureable but inconsistent geometric features, and exhibit rapidly changing scales, their detection is one of the most challenging tasks in target detection and tracking systems.
[0003] Targets often navigate through dense clouds and ocean waves in harsh environments and complex scenes, making detection tasks more challenging. Early methods lacked accuracy and robustness in detecting small targets, leading to the development of numerous new detection methods. Generally, all methods used to address this problem can be categorized into two types based on their detection approach: single-frame-based pre-tracking detection and continuous-frame-based pre-tracking detection. In terms of implementation, they can be broadly classified into three categories: low-rank sparse matrix-based recovery methods, contrast-based metric methods, and data-driven machine learning methods.
[0004] Using nonlinear adaptive filters to simulate clutter time profiles to extract real target time profiles is insufficient and incomplete, as it only considers target information in the local time domain. Sometimes images contain significant amounts of cloud cover and chaotic ocean backgrounds, making effective target detection difficult using only local spatial contrast. A multi-frame high-dimensional structural tensor model considering target structural prior knowledge and background autocorrelation suffers from high computational costs, requiring substantial hardware resources for real-time operation, and its accuracy is not ideal.
[0005] Detection methods based on contrastive metrics and low-rank sparse matrix recovery have reached a bottleneck, with various improvements offering limited performance enhancements. Meanwhile, higher demands are being placed on the detection of small targets. With data-driven support, deep learning methods can outperform traditional methods in terms of false negative rate, false alarm rate, and robustness. Therefore, researching the application of deep learning methods in the field of small targets is essential. Summary of the Invention
[0006] The purpose of this invention is to provide an image target recognition method based on self-attention and low-rank decomposition, which can perform security verification on the target recognition results of the image target recognition data to ensure reliability and interpretability.
[0007] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:
[0008] This invention provides an image target recognition method based on self-attention and low-rank decomposition, the method comprising:
[0009] S1: Obtain an existing public dataset;
[0010] S2: Based on the existing public dataset, use the IPI algorithm to generate a preliminary set of pseudo-labels;
[0011] S3: Merge the pseudo-label set, the real label set, and the existing public dataset into a training set; wherein each data point in the existing public dataset has a real label;
[0012] S4: Train the neural network using the training set to obtain the trained neural network;
[0013] S5: Perform loss calculation on the trained neural network to obtain the converged neural network;
[0014] S6: Use the converged neural network to traverse the training set to obtain the target recognition results of the training set;
[0015] S7: Perform inverse IPI algorithm processing on the target recognition results of the training set to obtain the processing result;
[0016] S8: Divide the feasible region using the processing results and the training set;
[0017] S9: Based on the target recognition image to be identified, the target recognition result of the target recognition data to be identified is obtained using the converged neural network;
[0018] S10: Perform feasible region discrimination on the target recognition result of the target recognition image to be identified. If the target recognition result of the target recognition image to be identified is within the feasible region, then output the target recognition result of the target recognition image to be identified.
[0019] Optionally, the neural network includes, in sequence, a first convolutional layer, a first pooling layer, a first residual block, a second pooling layer, a second residual block, a third pooling layer, a third residual block, a fourth pooling layer, a fourth residual block, a fifth pooling layer, a fifth residual block, a semantic-level self-attention module, a first fusion block, a second fusion block, a third fusion block, a sixth residual block, an interpretable self-attention module, a fourth fusion block, and a fifth fusion block;
[0020] The outputs of the third residual block, the fourth residual block, and the fifth residual block are simultaneously used as inputs to the semantic-level self-attention module. The output of the semantic-level self-attention module is simultaneously used as inputs to the first fusion block, the second fusion block, and the third fusion block. The output of the first convolutional layer is simultaneously used as inputs to the first pooling layer and the fifth fusion block. The output of the first residual block is simultaneously used as inputs to the second pooling layer and the fourth fusion block. The output of the second residual block is simultaneously used as inputs to the third pooling layer and the interpretable self-attention module.
[0021] Alternatively, the output of the first convolutional layer is 256×256×32;
[0022] The outputs of the first pooling layer and the first residual block are both 128×128×48;
[0023] The outputs of the second pooling layer and the second residual block are both 64×64×48;
[0024] The outputs of the third pooling layer and the third residual block are both 32×32×64;
[0025] The outputs of the fourth pooling layer and the fourth residual block are both 16×16×64;
[0026] The outputs of the fifth pooling layer and the fifth residual block are both 8×8×96;
[0027] The output of the first fusion block is 8×8×96;
[0028] The output of the second fusion block is 16×16×64;
[0029] The output of the third fusion block is 32×32×64;
[0030] The output of the sixth residual block is 64×64×48;
[0031] The output of the interpretable self-attention module is 64×64×48;
[0032] The output of the fourth fusion block is 128×128×48;
[0033] The output of the fifth fusion block is 256×256×1.
[0034] Optionally, the interpretable self-attention module includes, in sequence, a second convolutional layer, a seventh residual block, a third convolutional layer, a first two-dimensional vector processing submodule, a gradient-free iteration submodule, a coefficient inference submodule, and a second two-dimensional vector processing submodule.
[0035] The first two-dimensional vector processing submodule is used to flatten the length and width dimensions of the output of the third convolutional layer to obtain the input matrix; transpose the input matrix and perform matrix multiplication on the transposed input matrix and the basis matrix, and then normalize it to obtain the coefficient matrix; and transmit the input matrix, the basis matrix and the coefficient matrix to the gradient-free iteration submodule.
[0036] The gradient-free iteration submodule includes a first coefficient inference block and a base inference block set sequentially and is used to update the base matrix and the coefficient matrix to obtain the updated base matrix and the updated coefficient matrix;
[0037] The coefficient inference submodule includes a second coefficient inference block, which is used to optimize and update the gradient of the updated basis matrix to obtain the final basis matrix.
[0038] The second two-dimensional vector processing submodule is used to transpose the updated coefficient matrix and perform matrix multiplication between the transposed coefficient matrix and the final basis matrix to obtain a low-yield matrix; the low-yield matrix is subtracted from the input matrix to obtain the final matrix as the output of the interpretable self-attention module.
[0039] Alternatively, S5 includes:
[0040] The loss function is obtained based on the weighted binary cross-entropy loss and the Dice loss;
[0041] The loss function is used to calculate the loss of the trained neural network, resulting in a converged neural network.
[0042] Alternatively, the loss function LOSS includes:
[0043]
[0044] Where, x i y represents the value of the i-th pixel in the output. i w represents the value of the i-th pixel on the label. i This represents the value of the i-th pixel in the weighted image provided by the pseudo-label, where i represents the i-th pixel and n represents the total number of pixels.
[0045] Alternatively, in step S10, the feasible region determination includes three inverse processes: inverse normalization, inverse meanization, and inverse reconstruction.
[0046] The inverse normalization is expressed as:
[0047]
[0048] Among them, pred NetThis represents the target recognition result of the image to be identified, pred IPI This represents the output prediction graph of the IPI algorithm, pred De-nomalize This represents the result after inverse normalization;
[0049] The inverse meanization is expressed as:
[0050]
[0051] in, Let n represent the i-th value of the matrix after inverse mean-normalization. i This represents the number of pixels in the target image that constitutes the i-th value. This represents the i-th value of the coefficient matrix generated by the IPI algorithm. This represents the value of the j-th pixel in the target recognition result of the image to be identified;
[0052] The inverse reconstruction is represented as:
[0053]
[0054] Where A represents the low-yield matrix, A0 represents the optimized low-yield matrix, N represents noise, and δ represents the minimum value.
[0055] Alternatively, in step S10, the feasible region discrimination further includes an optimization process, which is used to optimize the low-yield matrix after inverse reconstruction using an accelerated proximal gradient algorithm to remove noise information from the low-yield matrix.
[0056] The present invention has the following beneficial effects:
[0057] The presence of the neural network in this invention enables interpretability of image target recognition. Furthermore, by performing feasible domain discrimination on the target recognition results of the image target recognition data, security verification of the target recognition results of the image target recognition data can be performed to ensure reliability. Attached Figure Description
[0058] Figure 1 This is a flowchart of the image target recognition method based on self-attention and low-rank decomposition of the present invention;
[0059] Figure 2 This is a framework diagram of the image target recognition method based on self-attention and low-rank decomposition of the present invention;
[0060] Figure 3 This is a schematic diagram of the structure of an interpretable self-attention module. Detailed Implementation
[0061] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.
[0062] This invention provides an image target recognition method based on self-attention and low-rank decomposition, referencing... Figure 1 As shown, the image target recognition method based on self-attention and low-rank decomposition includes:
[0063] S1: Obtain an existing public dataset;
[0064] The existing publicly available datasets obtained in this invention consist of three datasets: two are single-frame datasets (image data units are time-discrete), and the other is a multi-frame dataset (image data units are time-continuous). Specific information about the three datasets is as follows:
[0065] (1) SIRST dataset. This dataset consists of 427 images, including two subsets of 320 and 427 images for training, validation, and testing. The images vary in size and have pixel-level annotations. It covers scenes such as nearshore, offshore, low-altitude, and high-altitude, with particularly rich scenes featuring complex cloud cover.
[0066] (2) IRSTD Dataset. This dataset consists of 1000 infrared images captured in the real world by infrared cameras, with pixel-level annotations and an image size of 512×512. It contains different types of small targets, such as drones, organisms, boats, and vehicles, captured from various locations at relatively long imaging distances. The dataset covers many different scenes with backgrounds including oceans, rivers, fields, mountains, cities, and clouds, with significant clutter and noise.
[0067] (3) SkyCup Dataset. This dataset is designed for low-altitude, low-flying aircraft target detection and tracking applications. Through field photography and data preparation, it provides a test dataset for algorithms targeting one or more fixed-wing UAVs. The dataset covers various backgrounds including the sky and ground, and includes 22 data segments, 30 flight paths, 16,177 image frames, and 16,944 targets. Each target corresponds to a labeled location, and each data segment corresponds to a labeled file. This dataset can provide fundamental data for research on low-target detection, precision guidance, and infrared target characteristics.
[0068] S2: Based on the existing public dataset, use the IPI algorithm to generate a preliminary set of pseudo-labels;
[0069] It should be noted that the IPI algorithm is an existing technology (for details, please refer to "Gao C, Meng D, Yang Y, et al. Infrared Patch-Image Model for Small TargetDetection in a Single Image[J].IEEE Transactions on Image Processing,2013,22(12):4996-5009."), and will not be elaborated upon in this invention.
[0070] Label smoothing assumes that mislabeled samples inevitably exist in the training samples, especially for one-hot labels. If negative samples are mislabeled, this leads to lost loss, affecting the model's prediction of negative samples. If positive samples are mislabeled, this leads to increased loss, causing overconfidence in predicting positive samples, ultimately leading to overfitting. Therefore, the purpose of label smoothing is to add smoothing noise to the labels to achieve regularization. This regularization reduces overfitting and increases generalization performance, inter-class distance, and class clustering. Since infrared weak target recognition does not require semantic segmentation but saliency segmentation, it is essentially a binary classification problem with typical one-hot labels. Therefore, the optimization process for neural networks targeting infrared weak targets also suffers from the aforementioned problems. However, instead of introducing random noise, it is better to introduce pseudo-noise that is closer to the true labels than random noise, or in other words, images reconstructed from sparse matrices output by optimization theory algorithms.
[0071] Pseudo-label iterative optimization is a commonly used technique in semi-supervised learning. It trains the model with more than half of the correct labels and less than half of the incorrect labels, and then uses the model output as new labels to iteratively train the model to further optimize it. This technique can reduce inter-class overlap and decrease inter-class density when it is impossible to improve the labeling accuracy.
[0072] refer to Figure 2 As shown, this pseudo-label set can weight and enhance the binary cross-entropy loss in the loss function, and continuously update the pseudo-label set with random proportions using the model's output graph when the network model reaches a bottleneck on the validation set until the network model stops when it reaches a bottleneck on the validation set again.
[0073] S3: Merge the pseudo-label set, the real label set, and the existing public dataset into a training set; wherein each data point in the existing public dataset has a real label;
[0074] S4: Train the neural network using the training set to obtain the trained neural network;
[0075] refer to Figure 2As shown, the neural network provided by the present invention includes: a first convolutional layer, a first pooling layer, a first residual block, a second pooling layer, a second residual block, a third pooling layer, a third residual block, a fourth pooling layer, a fourth residual block, a fifth pooling layer, a fifth residual block, a semantic-level self-attention module, a first fusion block, a second fusion block, a third fusion block, a sixth residual block, an interpretable self-attention module, a fourth fusion block, and a fifth fusion block, arranged sequentially.
[0076] The outputs of the third residual block, the fourth residual block, and the fifth residual block are simultaneously used as inputs to the semantic-level self-attention module. The output of the semantic-level self-attention module is simultaneously used as inputs to the first fusion block, the second fusion block, and the third fusion block. The output of the first convolutional layer is simultaneously used as inputs to the first pooling layer and the fifth fusion block. The output of the first residual block is simultaneously used as inputs to the second pooling layer and the fourth fusion block. The output of the second residual block is simultaneously used as inputs to the third pooling layer and the interpretable self-attention module.
[0077] Alternatively, the output of the first convolutional layer is 256×256×32;
[0078] The outputs of the first pooling layer and the first residual block are both 128×128×48;
[0079] The outputs of the second pooling layer and the second residual block are both 64×64×48;
[0080] The outputs of the third pooling layer and the third residual block are both 32×32×64;
[0081] The outputs of the fourth pooling layer and the fourth residual block are both 16×16×64;
[0082] The outputs of the fifth pooling layer and the fifth residual block are both 8×8×96;
[0083] The output of the first fusion block is 8×8×96;
[0084] The output of the second fusion block is 16×16×64;
[0085] The output of the third fusion block is 32×32×64;
[0086] The output of the sixth residual block is 64×64×48;
[0087] The output of the interpretable self-attention module is 64×64×48;
[0088] The output of the fourth fusion block is 128×128×48;
[0089] The output of the fifth fusion block is 256×256×1.
[0090] Since small targets in infrared images, after meeting the scale definition of the International Society for Optics and Photonics (ISO), mostly undergo the conversion from semantic-level information to pixel-level information at the 64x64 resolution level, certain constraints and operations must be applied at this level to ensure the interpretability of the entire network model. Therefore, this invention incorporates an interpretable self-attention module (Sparser) at the 64x64 resolution level, which possesses a sparse bias property. The Sparser module of this invention is designed with a core structure based on concept decomposition and non-negative matrix decomposition; therefore, it inherently possesses a low-rank bias property. This means that during the learning process, it has the ability, or bias, to extract and aggregate low-rank information from the information integrated through semantic-level self-attention. For general image recognition and segmentation, converting from low-rank bias attributes to sparse bias attributes is quite difficult. However, infrared weak target recognition tasks rely on a well-known moderate-strength assumption: adding low-rank terms to sparse terms can reconstruct the original image relatively completely. Therefore, after low-rank truncation from the core structure of the Sparser module, simply subtracting the input and output vectors can reverse the module's attributes from low-rank bias to sparse bias. Furthermore, since the data domain output by the semantic-level self-attention mechanism module requires some data transformation to reach the data domain from which the Sparser module can extract sparse terms, such as... Figure 3 As shown, an additional residual block is added before the Sparser module to learn this data mapping.
[0091] refer to Figure 3 As shown, the interpretable self-attention module includes, in sequence, a second convolutional layer, a seventh residual block, a third convolutional layer, a first two-dimensional vector processing submodule, a gradient-free iteration submodule, a coefficient inference submodule, and a second two-dimensional vector processing submodule.
[0092] The first two-dimensional vector processing submodule is used to flatten the length and width dimensions of the output of the third convolutional layer to obtain the input matrix; transpose the input matrix and perform matrix multiplication on the transposed input matrix and the basis matrix, and then normalize it to obtain the coefficient matrix; and transmit the input matrix, the basis matrix and the coefficient matrix to the gradient-free iteration submodule.
[0093] The gradient-free iteration submodule includes a first coefficient inference block and a base inference block set sequentially and is used to update the base matrix and the coefficient matrix to obtain the updated base matrix and the updated coefficient matrix;
[0094] The coefficient inference submodule includes a second coefficient inference block, which is used to optimize and update the gradient of the updated basis matrix to obtain the final basis matrix.
[0095] The second two-dimensional vector processing submodule is used to transpose the updated coefficient matrix and perform matrix multiplication between the transposed coefficient matrix and the final basis matrix to obtain a low-yield matrix; the low-yield matrix is subtracted from the input matrix to obtain the final matrix as the output of the interpretable self-attention module.
[0096] In other words, the actual input to the Sparser module is the 48x64x64 feature tensor from the underlying decoder, which is then mapped by the residual block and upsized to a 96x64x64 feature tensor. This feature vector, after being flattened in both dimensions, yields a 96x4096 input matrix as shown in the diagram above. The transposed input matrix is then multiplied by the 96x64 basis matrix to be optimized, followed by softmax to obtain a 4096x64 coefficient matrix. The input matrix, basis matrix, and coefficient matrix are then fed into an iterative inference block without gradient requirements. This block consists of a basis inference block and a sparse inference block. After feedforward computation by this block, the basis matrix and coefficient matrix are updated. Next, a coefficient inference block with gradient requirements outputs the input matrix, basis matrix, and coefficient matrix. It's worth noting that the input matrix remains unchanged throughout the process, while the coefficient and basis matrices change. Only the gradient of the final coefficient inference block requires optimization and updating of the basis matrix.
[0097] At this point, the coefficient matrix is transposed and multiplied with the basis matrix to obtain the low-rank matrix. The low-rank matrix is then subtracted from the input matrix to obtain the final output matrix of the Sparser module. This output matrix, after reconstruction, will restore the input 96x64x64 feature tensor. The feature tensor then undergoes a simple convolutional dimensionality reduction before participating in subsequent decoding operations.
[0098] It should be noted that the coefficient inference block primarily references the update of the coefficient matrix using nonnegative matrix decomposition. Specifically, the base matrix, after transpose, is multiplied by itself to obtain a 64x64 matrix. The coefficient matrix is then multiplied by the 64x64 matrix to obtain a 4096x64 matrix. The input matrix, after transpose, is multiplied by the base matrix to obtain another 4096x64 matrix. The coefficient matrix is then multiplied by this 4096x64 matrix, and then divided by the previous 4096x64 matrix to obtain the new coefficient matrix. The coefficient inference block operation process can be represented by the following formula:
[0099]
[0100] Among them, C ijX represents the element in row i and column j of a sparse matrix. T Let D denote the transpose of matrix X, where X represents the input matrix and D represents the basis matrix. T Let D be the transpose of matrix D, and C be the transpose of matrix D. Let i and j represent the row and column, respectively.
[0101] The basis inference block primarily updates the basis matrix by referencing nonnegative matrix decomposition. Specifically, the coefficient matrix, after transposing, is multiplied by itself to obtain a 64x64 matrix. Multiplying the coefficient matrix by the 64x64 matrix yields a 96x64 matrix. Multiplying the input matrix by the coefficient matrix results in another 96x64 matrix. The basis matrix is then multiplied by this 96x64 matrix and divided by the previous 96x64 matrix to obtain the output matrix, which is the new basis matrix. The basis inference block operation can be represented by the following formula:
[0102]
[0103] Among them, D ij Let X represent the element in row i and column j of the basis matrix, X represent the input matrix, C represent the coefficient matrix, and D represent the basis matrix. T This represents the transpose of C, where i and j represent the row and column, respectively.
[0104] S5: Perform loss calculation on the trained neural network to obtain the converged neural network;
[0105] Alternatively, S5 includes:
[0106] The loss function is obtained based on the weighted binary cross-entropy loss and the Dice loss;
[0107] The loss function is used to calculate the loss of the trained neural network, resulting in a converged neural network.
[0108] Based on this, the loss function LOSS of this invention is:
[0109]
[0110] Where, x i y represents the value of the i-th pixel in the output. i w represents the value of the i-th pixel on the label. i This represents the value of the i-th pixel in the weighted image provided by the pseudo-label, where i represents the i-th pixel and n represents the total number of pixels.
[0111] S6: Use the converged neural network to traverse the training set to obtain the target recognition results of the training set;
[0112] S7: Perform inverse IPI algorithm processing on the target recognition results of the training set to obtain the processing result;
[0113] S8: Divide the feasible region using the processing results and the training set;
[0114] When applied in real-world scenarios, if the model response encounters problems and relevant indicators fall outside the feasible region, effective early warnings can be issued. On one hand, directly using high-dimensional data such as feature maps within the neural network to establish a multi-dimensional coordinate system to generate the feasible region is computationally intensive. On the other hand, dimensionality reduction of these feature maps to construct a low-dimensional coordinate system can avoid algorithmic overhead, but these feature maps lack effective theoretical justification and interpretability. This invention argues that the IPI algorithm is highly reliable and efficient for modeling infrared weak targets, and it is itself a model with a solid mathematical foundation. Therefore, the kernel norm of the low-rank term and the zero norm of the sparse term are highly effective indicators. The key lies in inverse reconstruction to recover the sparse and low-rank matrices. The reliability verification algorithm is an effective re-optimization algorithm for inverse reconstruction of the sparse and low-rank matrices, which can be used for reliability verification. In summary, after the model training is completed, the output is inversely reconstructed to recover the sparse matrix. The sparse matrix is subtracted from the observation matrix reconstructed from the output image to obtain the initial low-rank matrix. The low-rank matrix is then further optimized. By repeating this process, the nuclear norm and zero norm of the low-rank sparse matrix trained on the training set and validation set are continuously calculated to generate the feasible region. Finally, in practical applications, the reliability verification process is completed by checking whether the model is within the feasible region.
[0115] S9: Based on the target recognition image to be identified, the target recognition result of the target recognition data to be identified is obtained using the converged neural network;
[0116] S10: Perform feasible region discrimination on the target recognition result of the target recognition image to be identified. If the target recognition result of the target recognition image to be identified is within the feasible region, then output the target recognition result of the target recognition image to be identified.
[0117] Alternatively, the feasible region determination includes three inverse processes: inverse normalization, inverse mean normalization, and inverse reconstruction.
[0118] The inverse normalization process primarily helps the network model's output prediction graph recover values from 0 to 255 to the range consistent with low-rank sparse models. Therefore, the inverse normalization is expressed as:
[0119]
[0120] Among them, pred Net This represents the target recognition result of the image to be identified, pred IPI This represents the output prediction graph of the IPI algorithm, predDe-nomalize This represents the result after inverse normalization;
[0121] The inverse mean and inverse reconstruction processes primarily help restore data from an image type to a matrix type composed of image patches. Since the value of a pixel in the predicted image may be derived by averaging the data from multiple points in multiple columns of the matrix, the inverse mean is expressed as:
[0122]
[0123] in, Let n represent the i-th value of the matrix after inverse mean-normalization. i This represents the number of pixels in the target image that constitutes the i-th value. This represents the i-th value of the coefficient matrix generated by the IPI algorithm. This represents the value of the j-th pixel in the target recognition result of the image to be identified;
[0124] The inverse reconstruction is represented as:
[0125]
[0126] Where A represents the low-yield matrix, A0 represents the optimized low-yield matrix, N represents noise, and δ represents the minimum value.
[0127] Alternatively, in step S10, the feasible region discrimination further includes an optimization process, which is used to optimize the low-yield matrix after inverse reconstruction using an accelerated proximal gradient algorithm to remove noise information from the low-yield matrix.
[0128] At this point, a complete sub-process of reliability verification has been completed. Finally, it is only necessary to repeat this sub-process on a reliable dataset (such as the training set and the validation set), and plot the corresponding feasible regions using the nuclear norm of the optimized low-rank matrix A_0 and the 0 norm of the sparse matrix E as the horizontal and vertical axes.
[0129] This invention uses the Kaiming initialization method to initialize the network model parameters to prevent gradient explosion or vanishing problems. The Adam optimizer is selected to optimize the network parameters, and the initial learning rate is set to 0.1, the momentum to 0.9, the weight decay to 0.0001, and the batch size to 8.
[0130] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for image target recognition based on self-attention and low-rank decomposition, characterized in that, The image target recognition method based on self-attention and low-rank decomposition includes: S1: Obtain an existing public dataset; S2: Based on the existing public dataset, use the IPI algorithm to generate a preliminary set of pseudo-labels; S3: Merge the pseudo-label set, the real label set, and the existing public dataset into a training set; wherein each data point in the existing public dataset has a real label; S4: Train the neural network using the training set to obtain the trained neural network; S5: Perform loss calculation on the trained neural network to obtain the converged neural network; S6: Use the converged neural network to traverse the training set to obtain the target recognition results of the training set; S7: Perform inverse IPI algorithm processing on the target recognition results of the training set to obtain the processing result; S8: Divide the feasible region using the processing results and the training set; S9: Based on the target recognition image to be identified, the target recognition result of the target recognition data to be identified is obtained using the converged neural network; S10: Perform feasible region discrimination on the target recognition result of the target image to be recognized. If the target recognition result of the target image to be recognized is within the feasible region, then output the target recognition result of the target image to be recognized. The neural network includes, in sequence, a first convolutional layer, a first pooling layer, a first residual block, a second pooling layer, a second residual block, a third pooling layer, a third residual block, a fourth pooling layer, a fourth residual block, a fifth pooling layer, a fifth residual block, a semantic-level self-attention module, a first fusion block, a second fusion block, a third fusion block, a sixth residual block, an interpretable self-attention module, a fourth fusion block, and a fifth fusion block; The outputs of the third residual block, the fourth residual block, and the fifth residual block are simultaneously used as inputs to the semantic-level self-attention module. The output of the semantic-level self-attention module is simultaneously used as inputs to the first fusion block, the second fusion block, and the third fusion block. The output of the first convolutional layer is simultaneously used as inputs to the first pooling layer and the fifth fusion block. The output of the first residual block is simultaneously used as inputs to the second pooling layer and the fourth fusion block. The output of the second residual block is simultaneously used as inputs to the third pooling layer and the interpretable self-attention module. The interpretable self-attention module includes, in sequence, a second convolutional layer, a seventh residual block, a third convolutional layer, a first two-dimensional vector processing submodule, a gradient-free iteration submodule, a coefficient inference submodule, and a second two-dimensional vector processing submodule. The first two-dimensional vector processing submodule is used to flatten the length and width dimensions of the output of the third convolutional layer to obtain the input matrix; transpose the input matrix and perform matrix multiplication on the transposed input matrix and the basis matrix, and then normalize it to obtain the coefficient matrix; and transmit the input matrix, the basis matrix and the coefficient matrix to the gradient-free iteration submodule. The gradient-free iteration submodule includes a first coefficient inference block and a base inference block set sequentially and is used to update the base matrix and the coefficient matrix to obtain the updated base matrix and the updated coefficient matrix; The coefficient inference submodule includes a second coefficient inference block, which is used to optimize and update the gradient of the updated basis matrix to obtain the final basis matrix. The second two-dimensional vector processing submodule is used to transpose the updated coefficient matrix and perform matrix multiplication between the transposed coefficient matrix and the final basis matrix to obtain a low-yield matrix; the low-yield matrix is subtracted from the input matrix to obtain the final matrix as the output of the interpretable self-attention module.
2. The image target recognition method based on self-attention and low-rank decomposition according to claim 1, characterized in that, The output of the first convolutional layer is 256×256×32; The outputs of the first pooling layer and the first residual block are both 128×128×48; The outputs of the second pooling layer and the second residual block are both 64×64×48; The outputs of the third pooling layer and the third residual block are both 32×32×64; The outputs of the fourth pooling layer and the fourth residual block are both 16×16×64; The outputs of the fifth pooling layer and the fifth residual block are both 8×8×96; The output of the first fusion block is 8×8×96; The output of the second fusion block is 16×16×64; The output of the third fusion block is 32×32×64; The output of the sixth residual block is 64×64×48; The output of the interpretable self-attention module is 64×64×48; The output of the fourth fusion block is 128×128×48; The output of the fifth fusion block is 256×256×1.
3. The image target recognition method based on self-attention and low-rank decomposition according to claim 1, characterized in that, S5 includes: The loss function is obtained based on the weighted binary cross-entropy loss and the Dice loss; The loss function is used to calculate the loss of the trained neural network, resulting in a converged neural network.
4. The image target recognition method based on self-attention and low-rank decomposition according to claim 3, characterized in that, The loss function LOSS include: in, Indicates the output of the first i The value of each pixel. Indicates the first label i The value of each pixel. This represents the weighted graph provided by the pseudo-label. i The value of each pixel. i Indicates the first i 1 pixel n This represents the total number of pixels.
5. The image target recognition method based on self-attention and low-rank decomposition according to claim 1, characterized in that, In step S10, the feasible region determination includes three inverse processes: inverse normalization, inverse mean normalization, and inverse reconstruction. The inverse normalization is expressed as: in, This represents the target recognition result of the image to be identified. This represents the output prediction graph of the IPI algorithm. This represents the result after inverse normalization; The inverse meanization is expressed as: in, The matrix after inverse mean-normalization is represented as the first... i One value, Indicates the formation of the first i The value refers to the number of pixels in the image to be identified. Denotes the first element of the coefficient matrix generated by the IPI algorithm. i One value, The first part represents the target recognition result of the image to be recognized. j The value of each pixel; The inverse reconstruction is represented as: in, A Represents a low-yield matrix. This represents the optimized low-yield matrix. N Indicates noise. This represents the minimum value.
6. The image target recognition method based on self-attention and low-rank decomposition according to claim 5, characterized in that, In step S10, the feasible region discrimination further includes an optimization process, which is used to optimize the low-yield matrix after inverse reconstruction using an accelerated proximal gradient algorithm to remove noise information from the low-yield matrix.