Infrared dim small target detection method and system based on conflict filtering and feature focusing

Abstract

This invention discloses an infrared weak target detection method and system based on conflict filtering and feature focusing. The method includes: extracting features from an input image using a backbone network; performing multi-scale feature fusion and conflict filtering on the extracted feature maps; then performing multi-scale feature fusion and conflict filtering on the feature maps obtained from the backbone network to obtain a final feature map; performing feature focusing on the final feature map; obtaining a fused feature image based on the feature maps obtained from multiple feature focusing; performing multi-scale feature fusion on the feature maps obtained from multi-scale feature fusion and conflict filtering, the final feature map, and the fused feature image using a feature pyramid to obtain three feature maps of different scales; and inputting the three feature maps of different scales into a detection head. This invention can effectively suppress background feature interference, enhance target feature representation, and achieve high-precision infrared weak target detection.

Description

Technical Field

[0001] This invention belongs to the field of digital image processing and computer vision technology, and particularly relates to an infrared weak target detection method and system based on conflict filtering and feature focusing. Background Technology

[0002] Infrared small target detection technology, due to its advantages such as strong anti-jamming capability, good concealment, and wide night vision range, has been widely used in military fields such as air defense, missile defense, precision guidance, and maritime surveillance. Many infrared applications, such as missile tracking systems, infrared search and track systems, and early warning systems, can not only detect and track targets emitting infrared radiation but also cleverly conceal themselves. Furthermore, real-world scenarios require these applications to accurately and promptly detect targets. However, in actual scenarios, infrared imaging distances are extremely long, and the scenes are complex and varied, resulting in low signal-to-noise ratios and contrasts in the generated infrared images. Targets are often unclear due to the loss of significant details and texture features. Therefore, detecting small targets in low signal-to-noise ratio and complex backgrounds is a very difficult and challenging task.

[0003] In traditional methods, most researchers design detection operators based on the differences in grayscale and structural features between the target and the background in infrared images to directly extract the target, such as Top-hat, wavelet transform, matched filtering, and two-dimensional least mean square. Although these methods have low time complexity and fast convergence speed, they rely heavily on manual design and require strong prior assumptions, resulting in poor robustness and generalization ability. When the target is in complex weather conditions, manually constructed features and fixed hyperparameters are difficult to handle drastic scene changes. In recent years, Convolutional Neural Networks (CNNs) for target detection have been gradually developed. To effectively improve the detection accuracy of weak infrared targets in complex and rapidly changing scenes, data-driven CNN models have been proposed. Compared with the aforementioned traditional methods, their performance is more significant. Compared with traditional methods, they have excellent learning and feature representation capabilities, and can continuously learn target features from samples and continuously adjust model parameters to achieve more accurate target detection and recognition.

[0004] However, the detection performance of most data-driven methods is also significantly affected by rapidly changing backgrounds because most network models are too limited. First, to obtain deeper semantic features of weak infrared targets, these models overlay excessive convolutional and pooling layers, burying the target within the deep network. Second, to fuse high-level semantics and low-level details, these models directly add existing multi-scale feature fusion modules, but ignore a large amount of conflicting information during the fusion of features at different scales, limiting the expression of multi-scale features. Finally, due to the inherent dispersion of features of weak targets, the detection head struggles to notice important target features beneficial for detection, potentially leading to model failure. Therefore, professional network calibration is essential to improve the detection performance of weak infrared targets. Summary of the Invention

[0005] The technical problem solved by this invention is to overcome the shortcomings of the prior art and provide an infrared weak target detection method and system based on conflict filtering and feature focusing, which can alleviate the phenomenon of background noise overwhelming the target and the sparse and scattered target feature texture, effectively suppress background feature interference, enhance target feature representation, and achieve high-precision infrared weak target detection.

[0006] The objective of this invention is achieved through the following technical solution: an infrared weak target detection method based on conflict filtering and feature focusing, comprising: using a backbone network to extract features from an input image; performing multi-scale feature fusion and conflict filtering on the extracted feature maps; then performing multi-scale feature fusion and conflict filtering on the feature maps after multi-scale feature fusion and conflict filtering and the feature maps extracted by the backbone network to obtain a final feature map; performing feature focusing on the final feature map to obtain multiple feature maps after feature focusing; obtaining a fused feature image based on the multiple feature maps after feature focusing; using a feature pyramid to perform multi-scale feature fusion on the feature maps after multi-scale feature fusion and conflict filtering, the final feature map, and the fused feature image to obtain three feature maps of different scales; and inputting the three feature maps of different scales into a detection head for target classification and location regression.

[0007] In the above-mentioned infrared weak target detection method based on collision filtering and feature focusing, the collision filtering includes a channel filtering branch and a spatial filtering branch.

[0008] In the aforementioned infrared weak target detection method based on conflict filtering and feature focusing, the channel filtering branch obtains adaptive weights α, β, and γ for the first, second, and third channels, representing the spatial information of the global image features, by compressing the aggregated feature map in the spatial dimension. α, β, and γ are defined as follows:

[0009] [α i ,β i ,γi ]=σ[AP(F)]

[0010] Where F is the feature map generated by the concatenation operation, σ and AP are the sigmoid function and average pooling, respectively, and α i ,β i ,γ i Let α, β, and γ represent the adaptive weight values ​​α, β, and γ of the input to the i-th layer, respectively.

[0011]

[0012] Among them, C1 i It is the feature map of the i-th layer obtained after the channel filtering branch. and Let X and Y represent the feature values ​​at (X,Y) positions sampled from layer 0, layer p, and layer q to layer i, respectively, where X is the horizontal coordinate, Y is the vertical coordinate, and i is the layer number, i = 1, 2, 3.

[0013] In the infrared weak target detection method based on conflict filtering and feature focusing described above, the spatial filtering branch uses Softmax to normalize the feature map of the channel direction to obtain the relative weights of different channels at the same position.

[0014] In the aforementioned infrared weak target detection method based on conflict filtering and feature focusing, feature focusing is performed on the last feature map to obtain multiple feature maps after feature focusing, including: using pointwise convolution to extract single-point features through channel interaction to obtain feature map X1 after feature focusing; using global average pooling to encode the position to be suppressed to suppress image background information to obtain feature map X2 after feature focusing; and using max pooling to encode the position to be focused to retain more texture information to obtain feature map X3 after feature focusing.

[0015] In the above-mentioned infrared weak target detection method based on conflict filtering and feature focusing, the fused feature image obtained from multiple feature maps after feature focusing includes: adding the feature maps X1, X2, and X3 after feature focusing, aggregating important features from local and global features, obtaining the weight corresponding to each position through an activation function, and multiplying the corresponding positions of the original feature images to obtain the fused feature image.

[0016] In the aforementioned infrared weak target detection method based on conflict filtering and feature focusing, the input of feature maps of three different scales into the detection head for target classification and location regression includes: inputting feature maps of three different scales into the detection head for target category probability prediction and bounding box regression; continuously updating model parameters through backpropagation of the loss function until the model converges, thus obtaining an effective infrared weak target detection classification and localization model; and using the trained infrared weak target detection classification and localization model to perform feature extraction, feature fusion and conflict filtering, feature focusing, target category prediction and bounding box regression on the infrared image to be identified, to determine the target category probability and target location contained in the infrared image.

[0017] An infrared weak target detection system based on conflict filtering and feature focusing includes: a first module for extracting features from an input image using a backbone network, performing multi-scale feature fusion and conflict filtering on the extracted feature maps, and then performing multi-scale feature fusion and conflict filtering on the feature maps after multi-scale feature fusion and conflict filtering with the feature maps extracted by the backbone network to obtain a final feature map; a second module for performing feature focusing on the final feature map to obtain multiple feature maps after feature focusing, and obtaining a fused feature image based on the multiple feature maps after feature focusing; a third module for performing multi-scale feature fusion on the feature maps after multi-scale feature fusion and conflict filtering, the final feature map, and the fused feature image using a feature pyramid to obtain three feature maps of different scales; and a fourth module for inputting the three feature maps of different scales into a detection head for target classification and location regression.

[0018] In the infrared weak target detection system based on collision filtering and feature focusing described above, the collision filtering includes a channel filtering branch and a spatial filtering branch.

[0019] In the aforementioned infrared weak target detection system based on conflict filtering and feature focusing, the channel filtering branch obtains adaptive weights α, β, and γ for the first, second, and third channels, representing the spatial information of the global image features, by compressing the aggregated feature map in the spatial dimension. α, β, and γ are defined as follows:

[0020] [α i ,β i ,γ i ]=σ[AP(F)]

[0021] Where F is the feature map generated by the concatenation operation, σ and AP are the sigmoid function and average pooling, respectively, and α i ,β i ,γ i Let α, β, and γ represent the adaptive weight values ​​α, β, and γ of the input to the i-th layer, respectively.

[0022]

[0023] Among them, C1 i It is the feature map of the i-th layer obtained after the channel filtering branch. and Let X and Y represent the feature values ​​at (X,Y) positions sampled from layer 0, layer p, and layer q to layer i, respectively, where X is the horizontal coordinate, Y is the vertical coordinate, and i is the layer number, i = 1, 2, 3.

[0024] Compared with the prior art, the present invention has the following advantages:

[0025] This invention fully considers the problems of small target size and sparse target texture features in infrared weak target detection. By performing two multi-scale feature fusions and conflict filtering on feature maps of different scales obtained from the backbone network, background clutter is effectively suppressed, alleviating the phenomenon of conflicting information between different layers overwhelming the target during multi-scale feature fusion. By performing feature focusing on the final output of the backbone network, the focus on important target features is enhanced. Feature focusing integrates point-directed channels of three-branch paths, which can effectively preserve and highlight the subtle details of infrared weak targets, helping the network achieve more accurate target detection and recognition. Attached Figure Description

[0026] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0027] Figure 1 This is a flowchart of the infrared weak target detection method based on collision filtering and feature focusing provided in the embodiments of the present invention;

[0028] Figure 2 This is a schematic diagram of the conflict filtering process provided in an embodiment of the present invention;

[0029] Figure 3 This is a schematic diagram of the feature focusing process provided in an embodiment of the present invention. Detailed Implementation

[0030] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0031] Figure 1 This is a flowchart of an infrared weak target detection method based on collision filtering and feature focusing provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps:

[0032] Step S1: Use the backbone network to extract features from the input image, perform multi-scale feature fusion and conflict filtering on the extracted feature map to obtain a feature map after multi-scale feature fusion and conflict filtering, and then perform multi-scale feature fusion and conflict filtering on the feature map extracted by the backbone network to obtain the last feature map.

[0033] Step S2: Perform feature focusing on the last feature map to obtain multiple feature maps after feature focusing, and obtain a fused feature image based on the multiple feature maps after feature focusing;

[0034] Step S3: Use the feature pyramid to perform multi-scale feature fusion on the feature map after multi-scale feature fusion and conflict filtering, the last feature map, and the fused feature image to obtain three feature maps of different scales.

[0035] Step S4: Input feature maps of three different scales into the detection head to perform target classification and location regression.

[0036] Specifically, such as Figure 1 As shown, it includes the following steps:

[0037] S1: Use a backbone network to extract features from the input image, perform multi-scale feature fusion and conflict filtering on the extracted feature maps, and then perform multi-scale feature fusion and conflict filtering on the feature maps after multi-scale feature fusion and conflict filtering with the feature maps extracted by the backbone network.

[0038] (1) Image feature extraction

[0039] The system reads the input image and uses a backbone network to extract features from it. The backbone network is a fully convolutional network that does not use any pooling layers. It consists of five residual block units, each composed of convolutional units and residual blocks of different scales to extract feature maps at different scales. Each convolutional unit is composed of a convolutional layer, a batch normalization layer, and a LeakyReLu operation. Each residual block unit first downsamples the feature map using a convolutional unit with a kernel size of 3x3 and a stride of 2; then it extracts features using two residual blocks to obtain the corresponding scale feature map. Each residual block consists of two convolutional units with a kernel size of 1x1 and a stride of 1. The first 1x1 convolutional unit halves the number of channels in the feature map, and the second 1x1 convolutional unit restores the number of channels. A residual edge is then connected to the input to complete the residual connection. Specifically, the input image of length w and width h first passes through a convolutional unit with a kernel size of 3x3, a stride of 1, and padding of 1, generating a feature map of length w, width h, and 32 channels. Then, it passes through five residual units sequentially, generating five sets of feature maps with 64, 128, 256, 512, and 1024 channels respectively. The first residual block unit generates a feature map of length w... Width Feature map L1, then L1 enters the second residual block unit to generate a length of Width Feature map L2, then L2 enters the third residual block unit to generate a length of Width Feature map L3, then L3 enters the fourth residual block unit to generate a length of Width Feature map L4, then L4 enters the fifth residual block unit to generate a length of Width Feature map L5. The backbone network parameters used in this invention are shown in Table 1. Only convolution and residual connection operations, batch normalization, and LeakyReLu operations are listed in the table and are ignored.

[0040] Table 1 Backbone Network Parameters

[0041]

[0042] (2) Multi-scale feature fusion and conflict filtering

[0043] This invention takes into account that shallow features have small receptive fields, preserving more detailed information about the target boundary, while deep features have large receptive fields, preserving more semantic information about the target. By processing features at different layers, the feature information of small targets is enhanced. Specifically, for the five sets of feature maps with 64, 128, 256, 512, and 1024 channels, three feature maps with 128, 256, and 512 channels are selected for multi-scale feature fusion and conflict filtering. For ease of representation, these are defined as L1, L2, and L3, respectively, and the size of the L2 feature map is defined as C×H×W. Typical multi-scale feature fusion networks select the third-layer feature map or deeper feature maps for multi-scale feature fusion. However, this invention selects the second, third, and fourth-layer feature maps. This is because infrared weak targets are different from large and medium targets in general scenes. Infrared weak targets have low intensity, low contrast, low signal-to-noise ratio, and sparse and dispersed texture features. The shallower feature maps contain a lot of feature information. Although the deeper feature maps also contain some feature information, the feature information of the shallower layers cannot be ignored. After all, infrared weak targets have very few features, and it is necessary to minimize the loss and waste of any usable features. The specific process is as follows: Figure 2 As shown.

[0044] The L1 map with 128 channels is downsampled, and its scale is halved to obtain a feature map F1 of size C×H×W.

[0045] F1 = Downsample(L1)

[0046] Downsample(.) represents downsampling.

[0047] Convolve the feature map L2 with 256 channels to reduce its channel count to half and then restore it to its original size to obtain the feature map F2 with size C×H×W.

[0048] F2 = Conv(L2)

[0049] Conv represents a convolutional unit, which includes a convolution with a kernel size of 1x1 and 512 channels and a convolution with a kernel size of 1x1 and 256 channels, enabling dimensionality increase and decrease of the number of channels and completing information exchange between channels.

[0050] The L3 feature map with 512 channels is upsampled, and its scale is doubled to obtain a feature map F3 of size C×H×W.

[0051] F3 = Upsample(L3)

[0052] Upsample(.) represents upsampling.

[0053] F1, F2, and F3 are concatenated along the channel dimension to obtain a concatenated feature map F. This feature map has rich detail and semantic information and has a size of 3C×H×W.

[0054] F = Concat(F) 1 ,F 2 ,F 3 )

[0055] Where Concat is a feature tensor spliced ​​along the axis.

[0056] The stitched feature map integrates feature information from three different scales. However, directly stitching feature maps of different scales results in a large amount of conflict and redundant information. This information weakens or obscures the already sparse infrared weak target features, leading to fusion failure. Therefore, it is necessary to filter conflict information in the stitched feature map to effectively express the fused feature information. Thus, conflict filtering is performed on the stitched feature map F of size 3C×H×W. The filtering branches include channel filtering and spatial filtering. The channel filtering branch obtains 1×1×1 channel adaptive weights α, β, and γ representing the global features of the image by compressing the aggregated feature map in the spatial dimension. α, β, and γ can be defined as follows:

[0057] [α i ,β i ,γ i ]=σ[AP(F)]

[0058] Here, F is the feature map generated through concatenation. σ and AP are the sigmoid function and average pooling, respectively, and α... i ,β i ,γ i , (i = 1, 2, 3) represent the adaptive weight values ​​α, β, and γ of the i-th layer input, respectively.

[0059]

[0060] C1 i It is the feature map of the i-th (i=1,2,3) layer obtained after the upper branch. and Let O, P, and Q represent the feature values ​​sampled from layers O, P, and Q to position (X, Y) in layer i, respectively, where O, P, and Q = 1, 2, and 3 represent the three input feature maps.

[0061] The spatial filtering branch uses Softmax to normalize the feature map of the channel direction, obtaining the relative weights θ of different channels at the same location. and μ, the θ, μ can be defined as follows:

[0062]

[0063] Where, θ i , and μ i Represent the relative weights θ and θ of different channels at the same position in the i-th layer, respectively. μ and C are feature inputs from different channels at the same location.

[0064] Use the weight θ at the corresponding position of the channel i , and μ i Multiply the i-th layer feature map by the feature value of the corresponding position of the channel in the layer (i = 1, 2, 3), and then sum the weights of the corresponding positions of the channels in the layer to obtain the feature map of the i-th layer after spatial filtering branch purification.

[0065]

[0066] Where (K,X,Y) represents the position (X,Y) in channel K. These represent the spatial attention weights of the Kth channel relative to the corresponding layer. and Let C1, C2, and C3 represent the feature values ​​of channel K sampled from layers o, p, and q to position (X, Y) in layer i, respectively, where o, p, q = 1, 2, 3, representing the three input feature maps. i This is the feature map of the i-th layer after spatial filtering and purification.

[0067] The multi-scale feature fusion conflict filtering output can be defined as:

[0068] C i =C1 i +C2 i

[0069] Where C i It is the feature map of the i-th layer after multi-scale feature fusion and conflict filtering.

[0070] In this invention, the first total output of this process is C. 2 The scale corresponding to the original image is The output of the second use of this process is C. ′2 The scale corresponding to the original image is This concludes Wanheng's first step of feature extraction, multi-scale feature fusion, and conflict filtering, yielding the backbone network output feature map and feature map C used for feature pyramid stage fusion. 2 and C ′2 .

[0071] S2: Focus on the last feature map extracted from the S1 backbone network and weight the more important feature information;

[0072] This invention considers the sparse and dispersed features of small infrared targets. Although feature extraction, multi-scale feature fusion, and conflict filtering in S1 have enhanced the feature representation of small targets and suppressed background clutter interference, their enhancement capabilities are limited and cannot highlight important target features. Therefore, this invention uses feature focusing to focus on important target features, guiding the network to learn towards more critical features. The specific implementation of the feature focusing flowchart is as follows: Figure 3 As shown, the input feature map of this process is the output of the backbone network in S1, that is, the length is Width The feature map has 1024 channels and is downsampled by 32 times. This process defines the feature map as X, and its scale as C×H×W. As shown in the structural diagram, this process includes three branches. It borrows the idea of ​​attention mechanisms, achieving adaptive feature enhancement after feature extraction by changing the size of the spatial pool, thus improving the feature representation capability of deep, important features. Specifically, this process combines max pooling and average pooling, using pointwise convolution as a local channel context aggregator. This makes the network lightweight while fully utilizing points at each spatial location for pointwise channel interaction and aggregating the channel feature context of each spatial location. It includes three branches.

[0073] Branch 1: As a local channel context aggregator, pointwise convolution uses only pointwise channel interactions at each spatial location to extract features at a single point, increasing the number of channels in the feature map and reducing the loss and computational complexity of pooling operations. The input feature map X, with a scale of C×H×W and C channels, is processed by a first pointwise convolution with a kernel size of 1x1, reducing the number of channels C to its original value. Using Batch Normalization (BN) and ReLU to implement batch normalization and activation functions, we obtain a size of The feature map, then the The feature map is then fed into a second pointwise convolution with a kernel size of 1x1. This pointwise convolution reduces the number of channels. Restored to C, batch normalization and activation functions are implemented using BN and ReLU to obtain a feature map X1 of size C×H×W, where X1 is represented as...

[0074] X1=β(PWConv2(δ(β(PWConv1(X)))))

[0075] Where δ and β represent rectified linear unit (ReLU), batch normalization (BN), and the kernel sizes of PWConv1 and PWConv2 are respectively... and Where r is the channel scaling ratio.

[0076] Branch 2: Use global average pooling to encode the locations that need to be suppressed, thus suppressing image background information. Input a feature map X of scale C×H×W with C channels. Perform global average pooling on it using a convolution with kernel size H×W to obtain a C×1×1 feature matrix. Then, use a pointwise convolution with kernel size 1x1 on the C×1×1 feature matrix to reduce the number of channels C back to its original value. Using Batch Normalization (BN) and ReLU to implement batch normalization and activation functions, we obtain a size of The feature matrix, then the The feature matrix is ​​fed into a second pointwise convolution with a kernel size of 1x1. This pointwise convolution reduces the number of channels. Restored to C, batch normalization and activation functions are implemented using BN and ReLU to obtain a feature matrix of size C×H×W. The sigmoid function is used to map the element values ​​of the C×H×W feature matrix to the range (0, 1). The weights of the C×H×W feature matrix mapped to the range (0, 1) are multiplied by the element values ​​at the corresponding positions of the input X to obtain the output feature map X2 of branch 2. X2 is represented as...

[0077]

[0078] Among them Y1 and These represent the global feature contexts obtained through global average pooling and element-wise multiplication, respectively. σ represents the sigmoid function, and the representations of β, δ, PWConv1, and PWConv2 are the same as described above.

[0079] Branch 3: Use global max pooling to encode the locations that need to be suppressed, thus suppressing image background information. Input a feature map X of scale C×H×W with C channels. Perform global average pooling on it using a convolution with kernel size H×W to obtain a C×1×1 feature matrix. Then, use a pointwise convolution with kernel size 1x1 on the C×1×1 feature matrix to reduce the number of channels C back to its original value. Using Batch Normalization (BN) and ReLU to implement batch normalization and activation functions, we obtain a size of The feature matrix, then the The feature matrix is ​​fed into a second pointwise convolution with a kernel size of 1x1. This pointwise convolution reduces the number of channels. Restored to C, batch normalization and activation functions are implemented using Batch Normalization and ReLU to obtain a feature matrix of size C×H×W. The sigmoid function is used to map the element values ​​of the C×H×W feature matrix to the range (0, 1). The weights of the C×H×W feature matrix mapped to the range (0, 1) are multiplied by the corresponding element values ​​of the input X to obtain the output feature map X3 of branch 3. X3 is represented as...

[0080]

[0081] Among them Y2 and These represent the global feature contexts obtained through global average pooling and element-wise multiplication, respectively. The representations of σ, β, δ, PWConv1, and PWConv2 are the same as described above.

[0082] Fusion: The three outputs are summed to obtain important feature maps in the aggregated local and global feature contexts. Then, an activation function is used to obtain the weights corresponding to each position. The corresponding positions of the original feature images are multiplied to obtain the fused feature image X'. Input feature maps X1, X2, and X3, each of size C×H×W, are summed along the channel dimension to obtain a 3C×H×W feature map. The Sigmoid activation function is used to map the element values ​​of the 3C×H×W feature map to the range (0, 1), i.e., normalization is performed to obtain weights. The average of the 3-channel weights at each position in the 3C×H×W matrix is ​​then multiplied by the corresponding weight at the input feature map X to obtain the total output X' of the feature focusing process. X' is represented as...

[0083]

[0084] Where X' represents the output of feature map X after this feature focusing process, and its size is suppressed to the input feature map X, both being C×H×W, corresponding to an infrared image with an input of h×w. The actual length of this feature map is... Width

[0085] S3: Use the feature pyramid to perform multi-scale feature fusion on the feature maps obtained in S1 and S2 to obtain three different scales of feature maps;

[0086] This invention achieves multi-scale feature fusion of the neck region through a feature pyramid. The input of the multi-scale feature fusion is S1, and the resulting value is... Size feature map C 2 C ′2 The size obtained from S2 is X'. First, the first output of the feature pyramid is Feature map D1; simultaneously, D1 is upsampled to obtain the scale. The feature map and the feature pyramid input size are Feature map C ′2 By concatenating the components and using convolutional blocks to reduce the dimensions to 256, we obtain a result of size [size missing]. The output feature map D2; simultaneously, D2 is upsampled to obtain a size of The feature map and the feature pyramid input size are Feature map C 2 By concatenating the components and reducing the dimension to 128 using convolutional blocks, a result of size [size missing] is obtained. The output feature map D3.

[0087] S4: Input the feature maps of the three different scales obtained in S3 into the detection head to perform target classification and location regression.

[0088] The three different scale feature maps were obtained in S3. Feature map D1 Feature map D2 and The feature map D3. The detection head contains three detection branches with the same structure. Each branch includes a convolutional layer with a kernel size of 3×3 and a stride of 1, a BN layer, a LeakyReLU layer, and a convolutional layer with a kernel size of 1×1 and a stride of 1. The first convolutional layer can double the number of input feature map channels to obtain... 1024 and Three feature maps are then convolved to obtain... and There are three output dimensions, where 18 can be split into 3×(4+1+1), where 3 represents that each corresponds to 3 anchor boxes, 4 represents the center coordinates xy and width and height wh of the anchor box, 1 represents the confidence score of whether the anchor box contains the target, and 1 represents the total number of categories in the dataset.

[0089] Experimental Design:

[0090] Experimental dataset selection:

[0091] This paper uses an infrared image dataset for detecting and tracking small aircraft targets against a ground / air background. Nine sequences were selected as training and testing samples, and their details are shown in Table 2 below.

[0092] Table 2 Dataset Details

[0093]

[0094] Experimental setup:

[0095] The parameter selection for all algorithms employs a 5-round crossover method. First, the dataset is divided into two parts: one for training and one for testing. Then, the training set is further divided into five parts (1, 2, 3, 4, 5). First, parts 1-4 are trained, and part 5 is validated. Then, parts 1-5 are trained, and part 4 is validated, for a total of five rounds. The best parameters are then selected and tested on the test set.

[0096] Comparison method:

[0097] This invention compares the proposed infrared small target detection model based on conflict filtering and feature focusing with several advanced infrared small target detection methods. For model-driven methods, this invention selects three representative methods: Top-hat based on background suppression, IPI based on position decomposition, and LCM based on saliency detection. The detailed hyperparameter settings involved in their experiments are shown in Table II, all of which were determined by exhaustive search of the training set of the dataset. For data-driven methods, this invention also compares three representative methods: MDvsFA, cGAN, ALCNet, and ACM.

[0098] Evaluation indicators:

[0099] To better measure the performance of the proposed model, the selection of evaluation metrics needs to consider multiple factors. First, commonly used evaluation metrics in traditional infrared small target detection include background suppression factor and information gain ratio, but these metrics are often infinitely large in most cases. Second, traditional methods model infrared small target detection as a segmentation process, trading target integrity for higher detection accuracy. This approach is not suitable for semantic segmentation in deep networks. Finally, since the detector proposed in this paper uses the YOLOv3 detection head that regresses target location, category, and confidence, this invention selects commonly used deep learning target detection evaluation metrics at the pixel level, including P (precision), R (recall), F1 score, and AP (Average Pprecision). Furthermore, to better compare with traditional methods, pixel-level evaluation metrics IoU and nIoU are added to evaluate the model's localization ability, with nIoU serving as a complement to IoU.

[0100] Experimental results:

[0101] Table 3 Experimental Results

[0102]

[0103] Since most samples have high background clutter intensity, target size and shape differentiation, the detection results and scores of each algorithm in Table 3 show that the detection capability of this invention is significantly improved compared with traditional methods; compared with depth methods, this invention truly achieves the best precision, recall, F1 metric and AP, proving the effectiveness of this invention.

[0104] This embodiment also provides an infrared weak target detection system based on conflict filtering and feature focusing. The system includes: a first module for extracting features from an input image using a backbone network, performing multi-scale feature fusion and conflict filtering on the extracted feature maps, and then performing multi-scale feature fusion and conflict filtering on the feature maps after multi-scale feature fusion and conflict filtering with the feature maps extracted by the backbone network to obtain a final feature map; a second module for performing feature focusing on the final feature map to obtain multiple feature maps after feature focusing, and obtaining a fused feature image based on the multiple feature maps after feature focusing; a third module for performing multi-scale feature fusion on the feature maps after multi-scale feature fusion and conflict filtering, the final feature map, and the fused feature image using a feature pyramid to obtain three feature maps of different scales; and a fourth module for inputting the three feature maps of different scales into a detection head for target classification and location regression.

[0105] This invention fully considers the problems of small target size and sparse target texture features in infrared weak target detection. By performing two multi-scale feature fusions and conflict filtering on feature maps of different scales obtained from the backbone network, background clutter is effectively suppressed, alleviating the phenomenon of conflicting information between different layers overwhelming the target during multi-scale feature fusion. By performing feature focusing on the final output of the backbone network, the focus on important target features is enhanced. Feature focusing integrates point-directed channels of three-branch paths, which can effectively preserve and highlight the subtle details of infrared weak targets, helping the network achieve more accurate target detection and recognition.

[0106] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

Claims

1. A method for detecting infrared dim small target based on conflict filtering and feature focusing, characterized in that include: The backbone network is used to extract features from the input image. The extracted feature maps are then subjected to multi-scale feature fusion and conflict filtering. Finally, the feature maps after multi-scale feature fusion and conflict filtering are combined with the feature maps extracted by the backbone network to obtain the last feature map. The last feature map is subjected to feature focusing to obtain multiple feature maps after feature focusing. A fused feature image is obtained based on the multiple feature maps after feature focusing. The feature pyramid is used to perform multi-scale feature fusion on the feature map after multi-scale feature fusion and conflict filtering, the last feature map and the fused feature image to obtain three different scales of feature maps. Three feature maps of different scales are input into the detection head for target classification and location regression. The conflict filtering includes a channel filtering branch and a spatial filtering branch; The channel filtering branch obtains adaptive weights for the first, second, and third channels, representing the spatial information of the global image features, by compressing the aggregated feature map in the spatial dimension. The Defined as: in, It is a feature map generated through a splicing operation. and They are respectively Functional and average pooling, They represent the first Adaptive weight values ​​of layer input ; ； in, It is the first branch obtained after channel filtering. Layer feature mapping, , and They represent from the first Layer, First Layer, First Layer sampling to the first Layer Location feature value The x-axis is... The vertical axis is , The floor number is used for numbering, where, ; The spatial filtering branch uses... The feature map of the channel direction is normalized to obtain the relative weights of different channels at the same position; Feature focusing on the last feature map yields multiple feature maps after feature focusing, including: Pointwise convolution is used to extract single-point features through channel interaction, resulting in a feature map after feature focusing. Global average pooling is used to encode the locations that need to be suppressed, suppressing background information in the image and obtaining a feature map after feature focusing. Max pooling is used to encode the areas that need to be focused, preserving more texture information and resulting in a feature map after feature focusing. ; The fused feature image obtained from the feature maps after focusing on multiple features includes: Feature map after feature focusing Feature map after feature focusing and feature maps after feature focusing The important features in the local and global features are added together, and then the weights corresponding to each position are obtained through the activation function. The positions corresponding to the original feature images are multiplied together to obtain the fused feature image.

2. The infrared weak target detection method based on collision filtering and feature focusing according to claim 1, characterized in that: Three different scale feature maps are input into the detection head for target classification and location regression, including: Three feature maps of different scales are input into the detection head to perform target category probability prediction and bounding box regression. The model parameters are continuously updated through backpropagation of the loss function until the model converges, resulting in an effective infrared weak target detection, classification, and localization model. The trained infrared weak target detection, classification, and localization model is used to perform feature extraction, feature fusion and conflict filtering, feature focusing, target category prediction, and bounding box regression on the infrared image to be identified, to determine the target category probability and target location contained in the infrared image.

3. An infrared weak target detection system based on collision filtering and feature focusing, characterized in that... include: The first module is used to extract features from the input image using a backbone network, perform multi-scale feature fusion and conflict filtering on the extracted feature map, and then perform multi-scale feature fusion and conflict filtering on the feature map after multi-scale feature fusion and conflict filtering with the feature map extracted by the backbone network to obtain the final feature map. The second module is used to perform feature focusing on the last feature map to obtain multiple feature maps after feature focusing, and to obtain a fused feature image based on the multiple feature maps after feature focusing. The third module is used to perform multi-scale feature fusion on the feature map after multi-scale feature fusion and conflict filtering, the last feature map and the fused feature image using the feature pyramid, to obtain three different scale feature maps. The fourth module is used to input feature maps of three different scales into the detection head for target classification and location regression; The conflict filtering includes a channel filtering branch and a spatial filtering branch; The channel filtering branch obtains adaptive weights for the first, second, and third channels, representing the spatial information of the global image features, by compressing the aggregated feature map in the spatial dimension. The Defined as: in, It is a feature map generated through a splicing operation. and They are respectively Functional and average pooling, They represent the first Adaptive weight values ​​of layer input ; ； in, It is the first branch obtained after channel filtering. Layer feature mapping, , and They represent from the first Layer, First Layer, First Layer sampling to the first Layer Location feature value The x-axis is... The vertical axis is , The floor number is used for numbering, where, ; The spatial filtering branch uses... The feature map of the channel direction is normalized to obtain the relative weights of different channels at the same position; Feature focusing on the last feature map yields multiple feature maps after feature focusing, including: Pointwise convolution is used to extract single-point features through channel interaction, resulting in a feature map after feature focusing. Global average pooling is used to encode the locations that need to be suppressed, suppressing background information in the image and obtaining a feature map after feature focusing. Max pooling is used to encode the areas that need to be focused, preserving more texture information and resulting in a feature map after feature focusing. ; The fused feature image obtained from the feature maps after focusing on multiple features includes: Feature map after feature focusing Feature map after feature focusing and feature maps after feature focusing The important features in the local and global features are added together, and then the weights corresponding to each position are obtained through the activation function. The positions corresponding to the original feature images are multiplied together to obtain the fused feature image.

Images