Infrared weak and small target tracking method based on attention guidance proposal comparison

Abstract

The invention belongs to the field of target tracking, and particularly relates to an infrared weak and small target tracking method based on attention guidance proposal comparison, which comprises the following steps of: 1, preparing an infrared weak and small target tracking data set; 2, constructing a network model, wherein the model is composed of a feature extraction module, a hierarchical cross attention feature enhancement module, a global search proposal generation module, a comparison enhancement perception module and a dynamic template updating module; according to the method, cross-feature and cross-hierarchy information association is realized through a hierarchy cross attention module, and dependency between features is enhanced; a candidate proposal is generated by the global search module and then input into the contrast enhancement perception module, so that the gap between positive and negative samples is effectively enlarged, and the positive samples are closer to a real target; and finally, the output of the comparison module, the template features and the old dynamic template features are integrated through a dynamic template updating module to complete template updating, so that the robustness and the accuracy in the tracking process are improved, and the method is suitable for stable tracking of the infrared weak and small target under the complex background.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target tracking, in particular to an infrared dim small target tracking method based on attention guided proposal comparison. BACKGROUND

[0002] Infrared dim small target tracking technology has important application value in the fields of military early warning, reconnaissance and precision strike. However, due to the development of infrared imaging devices, thermal infrared images generally have low resolution and poor signal-to-noise ratio. In addition, factors such as long imaging distance, atmospheric disturbance and target motion interfere with the image, and the target in the image often appears as a small spot with unclear outline and missing texture, with extremely limited feature information. This brings three challenges: first, it is difficult to model the target features, and the stability is poor; second, it is easy to lose the target under the interference of occlusion, light change, motion blur, deformation and scale change; third, the adjacent noise interference is serious, which easily leads to false tracking. In recent years, deep learning-based methods have been widely used in infrared small target tracking, among which template feature matching methods perform well, but in the case of complex background and many false target interferences, continuous and stable tracking is still insufficient. Therefore, it is urgent to propose a more efficient and robust template matching and calibration mechanism to improve the matching accuracy of the target in consecutive frames and meet the application requirements of actual infrared dim small target tracking scenes.

[0003] The Chinese patent authorization number is "CN116630373B", and the name is "Infrared dim small target tracking method based on style re-calibration and improved twin network". The method takes the first frame image of the sequence as the template, and the Tth frame except the first frame as the detection image. Side window filtering is used for edge-preserving denoising, the processed image is input into the twin network to extract the feature map, and the feature fusion is performed through the SRM module multiplied by multiple terms to obtain the fusion feature map of the template frame and the detection frame respectively. The anchor with a width-height ratio of 1 is generated by inputting the RPN module to realize the position and size prediction of the infrared dim small target. The method has significant effect in suppressing the noise of the infrared image, but when the target deforms, the scale changes or similar targets exist, the matching accuracy and anti-interference ability are limited, which affects the stable tracking performance in complex scenes. SUMMARY

[0004] (I) Technical problems to be solved

[0005] In view of the deficiencies in the prior art, the present application provides an infrared dim small target tracking method based on attention guided proposal comparison, which solves the problem that the existing infrared dim small target tracking method is easily disturbed by the environment and causes the target to be lost.

[0006] (II) Technical solutions

[0007] In order to achieve the above purpose, the present application specifically adopts the following technical solutions:

[0008] An infrared dim small target tracking method based on attention guide proposal contrast, comprising the following steps:

[0009] Step 1, preparing a data set: the data set selected by the model is an infrared image dim small aircraft target detection and tracking data set in a ground / air background, the data is divided into a training set and a test set in proportion, wherein the training set is used for model training, and the test set data is used for model evaluation;

[0010] Step 2, constructing a network model: the entire infrared dim small target tracking network based on attention guide proposal contrast is composed of a feature extraction module, a hierarchical cross-attention feature enhancement module, a global search proposal generation module, a contrast enhancement perception module and a dynamic template updating module. The feature extraction module is composed of four convolutional blocks to realize the functions of down-sampling and feature extraction, the hierarchical cross-attention feature enhancement module is composed of a reorganization channel attention branch and a hierarchical cross-attention branch, the global search region proposal module is composed of a global feature correlation calculation part and a region proposal part, the contrast enhancement perception module includes a region similarity evaluation branch, a regression branch and a classification branch, and the dynamic template updating module realizes the template updating function by cross-correlation calculation and a multi-layer perception machine;

[0011] Step 3, designing a loss function and determining an evaluation index; selecting a loss function suitable for increasing the difference between positive and negative samples, target classification and coordinate regression for combination, using only binary cross-entropy loss L BCE and coordinate regression L smooth-L1 loss in the global search region proposal module, and the loss in the contrast enhancement perception module is composed of binary cross-entropy loss L BCE , coordinate regression loss L smooth-L1 and region difference loss function L con , so that the loss of the network reaches the best in the training process, and the performance evaluation in the target classification accuracy, coordinate prediction accuracy and prediction region accuracy is selected as the standard for selecting the best model.

[0012] Step 4, training and testing the network model: input the training set of the data set described in step 1 into the network for training, test each round of model, evaluate the obtained model under three indexes, and determine an end-to-end infrared dim small target image tracking model with good robustness.

[0013] Further, the data set selected by the model in step 1 is an infrared image dim small aircraft target detection and tracking data set in a ground / air background, only video sequences meeting the task requirements are selected here, the videos are divided into a training set and a test set in proportion, wherein the training set is used for model training, and the test set is used for model evaluation.

[0014] Further, the feature extraction module in step 2 comprises four feature extraction sub-blocks, wherein the first and second sub-blocks have the same structure, only the convolution kernel size of the convolution layer is different, the structure and convolution kernel size of the first and second sub-blocks are the same, and the first and second sub-blocks are composed of ordinary 2D convolution, batch normalization, max pooling layer and linear rectifier function; the third and fourth sub-blocks are composed of ordinary 2D convolution, batch normalization and linear rectifier function;

[0015] Further, the hierarchical cross-attention feature enhancement module in step 2 comprises a reorganization channel attention branch and a hierarchical cross-attention branch, the reorganization channel attention branch realizes cross-channel information flow and provides cross-feature channel information flow for the hierarchical cross-attention branch, the branch is composed of global average pooling, global maximum pooling, S activation function, multi-layer perception and channel splicing operation, wherein the multi-layer perception comprises multiple 1x1 convolution and linear rectifier function; the hierarchical cross-attention branch is composed of four 3x3 convolution layers, 1x1 convolution, layer normalization and normalized exponential function, the branch realizes cross-attention between cross-features by introducing the channel weight of the reorganization channel attention branch, improves the dependence between the two branches, and enhances the shared relevant information feature extraction.

[0016] Further, the global search proposal generation module in step 2 realizes the generation of positive and negative proposals, in which the pixel correlation is realized by pixel-by-pixel multiplication and multiple 1x1 convolution, and the ROIAlign and TOP-K algorithms are used to realize the proposal generation task, thereby providing the required information for the contrast enhancement perception module.

[0017] Further, the contrast enhancement perception module in step 2 comprises a contrast learning branch and a classification and coordinate regression branch, after the features of the two branches are preliminarily processed by 3x3 convolution, they are processed by the two branches respectively, in the region similarity evaluation branch, the similarity information between the target and the candidate proposal is calculated by feature segmentation, multi-layer perception, batch normalization, cosine similarity calculation and normalized exponential function; in the classification and coordinate regression branch, the two kinds of features are multiplied and integrated by global average pooling layer and full connection operation, and then the target classification probability and coordinate information are obtained after two parallel full connections, wherein the target classification probability is obtained by multiplying the classification branch and the contrast learning branch.

[0018] Further, the dynamic template update module in step 2 calculates and reorganizes the correlation between the template feature, the predicted feature and the dynamic template feature to obtain a new dynamic template feature for matching, in which the features are batch normalized, then matrix multiplication, normalized exponential function, multi-layer perception and residual connection are performed to obtain a new dynamic template feature for the contrast learning module, and the dynamic template feature is updated.

[0019] Further, the step 3 global search proposal generation module loss function design is realized by binary cross entropy loss and coordinate regression loss to constrain the reasonable generation of candidate proposal region; in the contrast enhancement perception module, the region difference loss function, the binary cross entropy loss and the coordinate regression loss are used to realize the function of expanding the difference between positive and negative samples and the accurate tracking of the target.

[0020] (Three) beneficial effects

[0021] Compared with the prior art, the present application provides an infrared dim small target tracking method based on attention guided proposal contrast, which has the following beneficial effects:

[0022] The present application designs a new infrared dim small target tracking technology combining region contrast and dynamic template updating, uses a hierarchical cross attention feature enhancement module to construct deep association between template features and search features, and realizes the function of accurate screening of candidate proposals through region contrast dynamic template updating, providing a new tracking framework for infrared dim small target tracking in complex background.

[0023] The present application designs a hierarchical cross attention feature enhancement module to enhance the correlation between features, introduces cross-level channel information in different branches to improve the information association between the two branches, further enhances the discriminability of feature information, so that the network can better adapt to extreme situations such as target deformation and occlusion, and improves the tracking accuracy by 1.36%.

[0024] The present application proposes a contrast enhancement perception module to improve the difference between positive and negative samples, which reduces the influence of interference on the classification ability of the network under the constraint of the region difference loss function, further improves the discrimination ability of the tracking network to interference features, improves the discrimination ability to target instances, and improves the tracking success rate by 1.12%.

[0025] The present application proposes a dynamic template updating module for target updating between adjacent frames, which obtains real-time dynamic template by calculating the correlation between template features, predicted features and dynamic template features, more flexibly fuses historical discriminative features and recent predicted features, and improves the correlation between template features and search features. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 The present application provides an infrared dim small target tracking flowchart based on attention guided proposal contrast;

[0027] Figure 2This is the structure diagram of the infrared dim target tracking based on attention-guided proposal comparison of the present invention;

[0028] Figure 3 This is a structural diagram of the feature extraction module of the present invention;

[0029] Figure 4 This is a structural diagram of a feature extraction sub-block in the feature extraction module of the present invention;

[0030] Figure 5 This is the structure diagram of the reorganized channel attention branch of the present invention;

[0031] Figure 6 This is a hierarchical cross-attention branch structure diagram of the present invention;

[0032] Figure 7 Generate a module structure diagram for the global search proposal of the present invention;

[0033] Figure 8 This is a structural diagram of the contrast enhancement perception module of the present invention;

[0034] Figure 9 This is a structural diagram of the dynamic template update module of the present invention. DETAILED DESCRIPTION

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0036] Example

[0037] like Figures 1-9 As shown, an embodiment of the present invention provides an infrared dim small target tracking method based on attention-guided proposal comparison, which specifically includes the following steps:

[0038] Step 1, prepare the data set: The dataset for detecting and tracking small aircraft targets in infrared images under ground / air background contains a total of 22 tracking sequences. However, since the targets of the first four tracking sequences are too large or multiple targets, only the last 18 tracking sequences are selected as training and test sets. The target object of this dataset is fixed-wing drones in the air. Its scenes include complex situations such as single target sky background, single target complex ground background, target moving from far to near, target moving from near to far, target leaving the field of view, and target returning to the field of view.

[0039] Step 2: Build a network model; Figure 2As shown, the infrared dim small target tracking network framework based on attention guidance proposal comparison is composed of a feature extraction module, a hierarchical cross attention feature enhancement module, a global search proposal generation module, a comparison enhancement perception module and a dynamic template update module. The feature extraction module is composed of four convolution blocks, the hierarchical cross attention feature enhancement module is composed of a recombination channel attention branch and a hierarchical cross attention branch, the global search proposal generation module is composed of a global feature correlation calculation part and a region proposal part, the comparison enhancement perception module includes a region comparison branch, a regression branch and a classification branch, and the dynamic template update module realizes the template update function by template correlation calculation and a multi-layer perception machine.

[0040] The four convolution blocks included in the feature extraction module share parameter weights when extracting the template image and the search image, wherein the template data is the first frame image of the sequence, and the search frame is other images except the first frame image of the sequence, such as Figure 3 As shown, the feature extraction subblock one and the feature extraction subblock two are the same in structure, and the feature extraction subblock three and the feature extraction subblock four are the same in structure, wherein the feature extraction subblock one and the feature extraction subblock two are as shown in Figure 4 (A), and the feature extraction subblock three and the feature extraction subblock four are as shown in Figure 4 (B). In the feature extraction subblock one, 11x11 convolution layers with a step of 2, batch normalization, maximum pooling and linear rectification functions are sequentially passed through, so that the channel number is adjusted from 3 channels to 96 channels of search features Ser1 and template features Tem1. The features output by the feature extraction subblock one are passed through 5x5 convolution layers with a step of 1, batch normalization, maximum pooling and linear rectification functions in the feature extraction subblock two, so that the channel number is adjusted from 96 channels to 256 channels of search features Ser2 and template features Tem2. Next, in the feature extraction subblock three, 3x3 convolution layers with a step of 1, batch normalization and linear rectification functions are passed through, so that the channel number is adjusted from 256 channels to 384 channels of search features Set3 and template features Tem3. Finally, in the convolution block three, 3x3 convolution layers with a step of 1, batch normalization and linear rectification functions are passed through, so that the channel number is adjusted from 384 channels to 256 channels of search features Ser4 and template features Tem4. Finally, (256, 24, 24) search features Ser4 and (256, 8, 8) template features Tem4 are obtained.

[0041] F1 = ReLU(Max(BN(Conv 11 (Pic))

[0042] F2 = ReLU(Max(BN(Conv5(F1)))

[0043] F3 = ReLU(BN(Conv3(F2)))

[0044] F4 = ReLU(BN(Conv3(F3)))

[0045] where Pic is the input image, F i is the output of the i-th convolutional block, ReLU is a linear activation function, BN is batch normalization, Conv 11 , Conv5 and Conv3 are 11x11 convolutional layers with a stride of 2, 5x5 convolutional layers with a stride of 1 and 3x3 convolutional layers with a stride of 1, respectively.

[0046] The reorganized channel attention branch and the hierarchical cross-attention branch in the hierarchical cross-attention feature enhancement module are composed of two branches and share weights, as shown in the reorganized channel attention branch. Figure 5 The first branch uniformly divides the input features by channel dimension to obtain C i , where i ∈ [1, 4], and each group of C i is uniformly divided again to obtain C ij , where j ∈ [1, 4], and the features C i1 , C i2 , C i3 and C i4 at the same position of different groups are concatenated, i.e., four groups of reorganized features RC′ i are obtained, where i ∈ [1, 4], and the four groups of reorganized features are concatenated again into complete features RC, which are input into global average pooling and multilayer perceptron to obtain reorganized channel features CA re In the second branch, the input features are respectively subjected to global average pooling and global maximum pooling, and then input into a multilayer perceptron to obtain low-frequency channel features CA gap and high-frequency channel features CA max The above three groups of features are added and subjected to S activation function to obtain hierarchical enhanced channel attention weights CA enhance The calculation formula of the reorganized channel attention module is as follows:

[0047] C1, C2, C3, C4 = Split(F4)

[0048] RC′1 = concat[C 11 , C 21 , C 31 , C 41 ]

[0049] RC = concat[RC′1, RC′2, RC′3, RC′4]

[0050] CA re = MLP(RC)

[0051] CA gap = MLP(GAP(F4))

[0052] CA max = MLP(GMP(F4))

[0053] CA enhance = Sigmoid(add(CA re , CA gap , CA max ))

[0054] where split is channel uniform division, concat is channel dimension splicing operation, GAP is global average pooling, GMP is global maximum pooling, MLP is multi-layer perceptron, add is pixel addition, and Sigmoid is S activation function.

[0055] The input feature 1 of the hierarchical cross attention branch is the feature of this branch, as shown in FIG. 4, the input feature 1 is first processed by batch normalization and three parallel convolution layers one to three, wherein the convolution layer one to three is a 3x3 deep convolution with a step of 1, and the template branch and the search branch will respectively obtain (256, 6, 6) and (256, 22, 22). After the calculation of the convolution layer one and the convolution layer two, the required Q feature matrix and K feature matrix are obtained. The feature calculated by the convolution layer three is multiplied by the output weight of the hierarchical cross attention module with the input feature 2, and then processed by batch normalization to obtain the V feature matrix containing the cross feature level enhancement information. After the matrix multiplication of the transpose of the Q feature matrix and the K feature matrix, the normalized exponential function is used for processing, and then the matrix multiplication is performed with the V feature matrix. After the correlation Att obtained by the Q feature matrix, the K feature matrix and the V feature matrix is processed by a 1x1 convolution, the output feature X of the 3x3 convolution with a step of 1 is multiplied by the input feature, and then processed by a multi-layer perceptron to obtain the hierarchical enhancement feature HEF. The calculation formula in the search branch is as follows: Figure 6 Q ser , K ser = BN(Conv3(Ser4))

[0056]

[0057]

[0058] X Ser = Conv3(Ser4)

[0059]

[0060] The calculation formula of the hierarchical cross attention branch in the template branch is as follows:

[0061] Q Tem , K Tem = BN(Conv3(Tem4)) ​

[0062]

[0063] X Tem =Conv3(Tem4)

[0064]

[0065] Where Conv3 is 3×3 convolution, Conv1 is 1×1 convolution, BN is batch normalization, is the weight of the reorganized channel attention module in the template branch, MLP is a multi-layer perceptron, and Softmax is a normalized exponential function.

[0066] The global search proposal generation module is as follows Figure 7 As shown, the input of this module has template features HEF Tem and search feature HEF Ser Together, the template feature HEF of the template feature branch Tem Use 6×6 convolution with a stride of 6 to obtain a feature map of (256, 1, 1) and compare this feature with the search feature HEF Ser Pixel-by-pixel multiplication is performed to obtain pixel-level correlation, and a 1×1 convolution with a step size of 1 is used to obtain a refined correlation map of (256, 22, 22). Two parallel 1×1 convolutions are used to obtain preliminary classification-related features (2, 22, 22) and preliminary regression-related features (4, 22, 22). The preliminary classification-related features F cls and the initial regression related features F reg In the input region proposal, the RoI Align algorithm and TOP-K algorithm are performed in the region module in sequence to obtain a feature map P consisting of 20 (256, 6, 6) features. i , where i∈[1, 20], the 20 P i List of candidate proposals proposal , the loss function used in this part is Loss First ,The calculation formula of the global search proposal generation module is as follows:

[0067] F cls =Conv1(Conv1(Conv6(HEF Tem )⊙HEF Ser ))

[0068] F Reg =Conv1(Conv1(Conv6(HEF Tem )⊙HEF Ser ))

[0069] List proposal =RoI Align[(TOP-K(Fcls )), F Reg ]

[0070] Where Conv1 and Conv6 are 1×1 convolution and 6×6 convolution with a stride of 6, respectively. ⊙ is the element-wise multiplication. The RoIAlign algorithm and TOP-K together form the region proposal function to realize the function of the candidate proposal list.

[0071] List the above candidate proposals proposal , template feature HEF Tem Dynamically update feature DTMs with legacy Last time Common input contrast enhancement perception module and the loss function used by this module is Loss Second ,like Figure 8 As shown, the template feature HEF Tem After concatenating with the dynamically updated feature DTM, a new template feature Tem is adjusted to (256, 6, 6) using 1×1 convolution. new and a list of candidate proposals proposal P in i Further feature extraction is performed through 3×3 convolution with a step size of 1 and a padding of 1. In the region similarity evaluation branch, the candidate proposal features are compared with the new template features Tem new According to the length and width dimensions, the input features are cut into 9 groups of (256, 2, 2) feature maps, and then multi-layer perceptron, batch normalization and flattening operations are performed to obtain the candidate region feature group Ser patch With the new template feature Tem patch , the corresponding position of Ser patch and Tem patch Calculate the cosine similarity S at the corresponding position mn , where m represents the candidate proposal list List proposal The mth candidate proposal in the , n represents the nth region Ser after the mth candidate proposal is divided patch or Tem patch , add up the cosine similarity values ​​of the 9 regions and take the average as the similarity probability Sim of the candidate proposal m , list the candidate proposals List proposal After the calculation is completed, the normalized exponential function is used to obtain the normalized probability Sim of each candidate proposal, which can be expressed as follows:

[0072] Tem new =Conv1(concat(HEF Tem , DTM))

[0073]

[0074] In the formula For padding 1 by 3x3 convolution, Conv1 is 1x1 convolution, BN is batch normalization, MLP is multi-layer perceptron, Softmax is normalized exponential function, Flatten is flattening operation, Split is feature segmentation operation.

[0075] In the classification and coordinate regression branch, in Figure 8 The new template feature Tem new and the P proposal in the candidate proposal list List i respectively are subjected to 3x3 convolution with a step of one and padding of one, then pixel multiplication is performed, global average pooling is performed, and a fully connected layer is performed to obtain an output H as (1024, 1, 1), the classification probability is obtained by multiplying the second fully connected layer with the normalized probability Sim of the candidate proposal, and the coordinate prediction is obtained by passing H through the second fully connected layer to obtain an output of 4 representing the coordinate point (X, Y) and the offset Δx and Δy. This module can be calculated using the following formula:

[0076] H = Liner(GAP(Tem new ⊙List proposal ))

[0077] P Cls = Liner(H) x Sim

[0078] x, y, Δx, Δy = Liner(H)

[0079] Proposal Hight = Max(P cls )

[0080] In the formula, Liner is a fully connected layer, and Proposal Hight is the highest probability prediction feature.

[0081] In the dynamic template update module, as shown in Figure 9 The module is used to realize dynamic updating of the template, and the template feature HEF Tem , the best prediction Proposal Hight , and the old dynamic template feature DTM Last time are used as input features. First, the best prediction Proposal Hight and the old dynamic template feature DF Last time are batch normalized, then pixel-by-pixel multiplication is performed, and a normalized exponential function is used to obtain a similarity probability. Then, the template feature HEF Tem is batch normalized, multiplied by the multi-layer perceptron, and then added to the dynamic feature DTM Last time to obtain the latest dynamic template feature DTM New . The formula is as follows:

[0082] DTM New = MLP(softmax(BN(Proposal High ) x BN(DTM Last time )) x BN(HEF Tem )) + HEF Tem

[0083] where MLP is a multi-layer perceptron, BN is batch normalization, and Softmax is a normalized exponential function.

[0084] Step 3 designs the loss function and determines the evaluation index to measure the difference between the predicted target and the true target and effectively evaluate the model performance, and then continuously optimize the network parameters to make more accurate predictions.

[0085] The loss function Loss First used by the global search proposal generation module is composed of binary cross-entropy loss L BCE and coordinate regression loss L smooth-L1 , as follows:

[0086] Loss First = L BCE + L smooth-L1

[0087] For the target category prediction accuracy, we use the binary cross-entropy loss:

[0088]

[0089] where yi is the probability of correct prediction, and y is the probability of incorrect prediction.

[0090] For ensuring the accuracy of target coordinate information, we use the coordinate regression loss:

[0091]

[0092] where Smooth L1 is the smooth-L1 function, defined as follows:

[0093]

[0094] The loss function used by the contrast learning module is Loss second composed of region difference function L con , binary cross-entropy loss L BCE , and coordinate regression loss L smooth-L1 , as follows:

[0095] Loss second = L con + LBCE +L smooth-L1

[0096] Binary Cross Entropy Loss L BCE and coordinate regression loss L smooth-L1 As described above.

[0097]

[0098] In the formula, Z is a dynamic update feature DTM, X + is the best prediction Proposal Hight , X - is a feature other than the best prediction Proposal Hight .

[0099] The obtained network model is evaluated under the evaluation of precision, success rate and state accuracy, and the performance of the model is comprehensively evaluated.

[0100] Step 4, training and evaluating the network model; in the training process, the used optimizer is an SGD random gradient descent optimizer, the momentum and weight decay are set to 0.9 and 1x10 -4 , the iteration number is 30, each pair of images in the training comes from the same sequence, the maximum distance between frames is set to 10, the obtained model is evaluated under three indexes, and an end-to-end infrared dim small target image tracking model with good robustness is determined.

[0101] The activation function used in the application is as follows:

[0102] (1) S activation function:

[0103]

[0104] (2) Normalized exponential function:

[0105]

[0106] (3) Linear rectifier function:

[0107]

[0108] The application constructs a new region contrast and dynamic template combined infrared dim small target tracking method based on attention guide proposal comparison, inputs the infrared image sequence into the model, and completes the infrared dim small target tracking task without other steps, so that the robustness of the task is greatly improved.

[0109] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for tracking small infrared targets based on attention-guided proposal comparison, characterized by: The method includes the following steps: Step 1: Prepare the dataset: Select a dataset of small aircraft target detection and tracking in infrared images under ground / air background and divide it into a training set and a test set; Step 2: Build a network model. The entire infrared small target detection network includes a feature extraction module, a hierarchical cross-attention feature enhancement module, a global search proposal generation module, a contrast enhancement perception module, and a dynamic template update module. The hierarchical cross-attention feature enhancement module includes a reorganized channel attention branch and a hierarchical cross-attention branch, wherein the reorganized channel attention branch reconstructs channel information by channel grouping, feature aggregation, and a multi-layer perceptron. This module is used to enhance cross-channel information flow and enhance channel information. The hierarchical cross-attention module uses convolution and similarity calculation operations, combined with the reorganized channel attention branch weights, to achieve feature correlation between cross-channels and cross-features, thereby enhancing the feature association between the template frame and the search frame. The contrast enhancement perception module includes two branches. In the regional contrast branch of contrastive learning, multiple sets of convolutions are used to achieve the fusion of dynamic template features and template features. The similarity of the candidate proposal is obtained through segmentation operation and normalized exponential function. The regression and classification branches realize the classification and coordinate information prediction of the target. The dynamic template update module calculates the correlation of the three input features and realizes feature fusion with the multi-layer perceptron, thereby dynamically updating the dynamic template in the dynamic template update module and serving as the input of the dynamic template feature in the contrast enhancement perception module; Step 3: Design a loss function and determine the evaluation metric. Determine the composite loss function for the global search proposal generation module and the contrast enhancement perception module so that the network composite loss value during training approaches the optimal value. Step 4: Train and evaluate the network model. Input the training set in the tracking dataset in step 1 into the network for training, and test each round of the model with the test set. Evaluate the resulting model under three indicators to determine whether a robust end-to-end infrared dim target image tracking model is achieved.

2. The infrared dim small target tracking method based on attention-guided proposal comparison according to claim 1 is characterized in that: The hierarchical cross-attention feature enhancement module in step 2 includes a reorganized channel attention branch and a hierarchical cross-attention branch, wherein the reorganized channel attention branch contains two branches in total, wherein the first branch is composed of channel segmentation, feature reorganization, a global average pooling layer and a multi-layer perceptron, and the second branch is composed of a parallel global average pooling layer and a multi-layer perceptron and a global maximum pooling and a multi-layer perceptron. The three sets of features are weighted and normalized by an exponential function to realize the flow of channel dimension and global channel information; in the hierarchical cross-attention branch, the feature matrix is ​​generated by three 3×3 convolutions and the reorganized channel attention branch is applied to the v feature matrix to provide the channel feature in another feature, and correlation calculation, 1×1 convolution, corresponding element multiplication, multi-layer perceptron and corresponding element addition are performed in sequence to realize the cross-feature information correlation with the other branch, thereby improving the deep dependence of the two branches on the channel feature and spatial feature.

3. The infrared dim small target tracking method based on attention-guided proposal comparison according to claim 1 is characterized in that: The contrast enhancement perception module in step 2 includes using 1×1 convolution and 3×3 convolution to realize the fusion of dynamic template features and template features. In the regional similarity evaluation branch, segmentation operation, multi-layer perceptron, normalized exponential function and cosine similarity are used to calculate the similarity of the candidate proposal. In the classification and regression branch, corresponding element multiplication, global average pooling layer and three full connections are used to realize the classification and coordinate information prediction of the target.

4. The infrared dim small target tracking method based on attention-guided proposal comparison according to claim 1 is characterized in that: In the step 2, the common template features, prediction features and dynamic template features input into the dynamic template update module are batch normalized through three parallel branches and similarity is calculated. The similarity is input into the multi-layer perceptron and then added element by element to the template features to obtain the new dynamic feature template update input dynamic template features, which are used as the dynamic template feature input in the contrastive learning module.

Images