Infrared weak and small target detection method based on multi-frame interactive perception

Abstract

The invention discloses an infrared weak and small target detection method based on multi-frame interactive perception, and the method comprises the steps: 1, carrying out the preprocessing of an infrared sequence image, and avoiding the overfitting in a training process; 2, constructing a time sequence feature extraction module for continuous multi-frame images; 3, constructing a feature interaction perception module with a U-shaped structure; 4, establishing a loss function between a result obtained after the infrared image of the current frame passes through the prediction module and the real binary image; and 5, training the network by using an infrared image, and detecting an infrared weak and small target by using the trained model. According to the method, the spatial-temporal information of the target is obtained by using multiple frames of images, and cross-level information fusion is realized, so that the accuracy of infrared weak and small target detection can be improved, and the capturing capability of a fast moving small target is particularly enhanced.

Description

Technical Field

[0001] The invention belongs to the field of computer vision, and specifically is an infrared dim small target detection method based on multi-frame interactive perception. Background Art

[0002] Infrared imaging is widely used in civil, military, industrial and remote sensing fields due to its strong anti-interference ability to changes in illumination. As a key technology in this field, the infrared small target (IRST) detection algorithm has attracted much attention from scholars at home and abroad. It aims to accurately locate small targets under various infrared backgrounds. However, due to the influence of the target radiation characteristics and the transmission path, small targets in infrared images often show low signal-to-noise ratio, small size and fast movement, making the detection task extremely challenging.

[0003] Traditional single-frame infrared small target detection (SIRST) methods, such as those based on multi-filter, sparse low-rank, and human visual system (HVS), have achieved results in specific fields, but they are highly dependent on manual feature design and prior knowledge. When faced with drastic changes in scenes, these traditional methods are difficult to adapt to new environments due to fixed hyperparameters, and their performance is greatly reduced. In contrast, deep learning-based methods have demonstrated powerful modeling capabilities and significantly improved performance by learning to extract highly discriminative features. However, these methods still face challenges in complex backgrounds, are sensitive to noise, and are limited in capturing transient features of fast-moving small targets.

[0004] Current multi-frame IRST detection methods usually use a single-frame detector to process multi-frame inputs, trying to retain dynamic information in the transformed single-frame image, but these methods are not designed for multi-frame scenarios, so the effective use of temporal information remains to be explored. In addition, with the rise of Vision Transformer (ViT), it is widely used in global image modeling by decomposing image patches and calculating correlations.

[0005] Although target detection methods based on deep learning have made significant progress, existing infrared dim target detection methods still have the following shortcomings:

[0006] 1. Small targets in infrared images often exhibit characteristics such as low signal-to-noise ratio, small size and fast movement. Existing single-frame methods face challenges in complex backgrounds, are sensitive to noise, and are limited in capturing the transient characteristics of fast-moving small targets.

[0007] 2. Existing multi-frame IRST detection methods are not designed for multi-frame scenarios, so the utilization of temporal information needs further exploration.

[0008] 3. Compared with the improved network structure, the current method only builds long-distance dependencies on a single feature map in the application of spatial self-attention (SA), rather than establishing contextual connections across levels. Summary of the invention

[0009] The present invention aims to solve the shortcomings of the above-mentioned prior art and proposes an infrared dim small target detection method based on multi-frame interactive perception, in order to capture the motion information of the target and realize cross-level information fusion, thereby suppressing background noise, retaining the edge contour features of infrared dim small targets, and enhancing the ability to capture fast-moving small targets.

[0010] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0011] The infrared dim small target detection method based on multi-frame interactive perception of the present invention is characterized in that it is carried out according to the following steps:

[0012] Step 1: preprocessing the infrared sequence images;

[0013] After obtaining the infrared image sequence and performing preprocessing such as resizing, cropping, horizontal and vertical flipping, the processed infrared image sequence is obtained, which is denoted as ;in, Represents the infrared image of the tnth frame in history. When n=0, let represents the infrared image of the current frame; k represents the number of frames in the infrared sequence image; let The true label is recorded as ;

[0014] Step 2: Construct a temporal feature extraction module for continuous multi-frame images, including: center difference convolution block, temporal attention module and residual convolution block;

[0015] Step 2.1: The central difference convolution block uses formula (1) to Calculate and get the initial feature F of the infrared image of the tnth frame t-n , thus obtaining the initial features of the infrared image sequence ; Among them, when n=0, F t Indicates the infrared image of the current frame The initial characteristics of

[0016] (1)

[0017] In formula (1), is an integer between [-1,1], Respectively represent the offset of the horizontal and vertical coordinates; express The pixel value at the center position (x, y) of express Middle position The pixel value at It's location The weights to be learned at is a hyperparameter, and θ ∈ [ 0 , 1 ] ;

[0018] Step 2.2: Temporal attention module Processing, the initial feature F of the current frame t and the initial feature F of the history frame t-1 t-1 After pairing, the temporal features of the historical frame t-1 are obtained , thus obtaining the temporal characteristics of the infrared image sequence ;

[0019] Step 2.3: The temporal features of the infrared image sequence and the initial feature F of the current frame t After splicing, it is input into the residual convolution block for processing to obtain the infrared image of the current frame Temporal and spatial characteristics of small mid-infrared targets st ;

[0020] Step 3: Construct a U-shaped feature interaction perception module, including: feature interaction perception module encoder, decoder and multi-scale semantic interaction Transformer block, and use it to st After processing, the interactive perception feature F of infrared weak targets is obtained A ;

[0021] Step 3.1: The feature interaction perception module encoder uses m groups of residual convolution blocks containing downsampling to generate F st Processing to obtain high-level features of the target ,in, represents the target high-level features output by the i-th residual convolution block, C i is the channel dimension of the output of the i-th group of residual convolution blocks;

[0022] Step 3.2: Use convolutions of different sizes to verify the high-level features of the target Perform patch embedding to obtain a series of embedding layer features ,in, represents the i-th embedding layer feature, and h represents The width of , w represents The height of ;

[0023] Step 3.3: Use the multi-scale semantic interaction Transformer block to embed the features of the layer Processing to obtain the interactive layer features ,in, represents the i-th interaction layer feature;

[0024] Step 3.4: Use the decoder to decode the interactive layer feature sequence Perform decoding operation to obtain the interactive perception feature F of infrared weak target A ;

[0025] Step 3.5: F A After being processed by residual convolution blocks and 1×1 convolutions, we get The infrared prediction results of t ;

[0026] Step 4: Use formula (10) to establish The total loss function L t ;

[0027] (10)

[0028] In formula (10), express The weighted binary cross entropy loss of ; express The maximum mean square error loss of express Dice loss; , are two weighting coefficients;

[0029] Step 5: Based on The network composed of the time series feature extraction module and the feature interaction perception module is trained using the gradient descent algorithm, and the total loss function L is calculated to update the network parameters. When the number of training iterations reaches the set number, the training is stopped to obtain the trained infrared dim small target detection model, which is used to realize the detection of infrared dim small targets in complex backgrounds.

[0030] The infrared dim small target detection method based on multi-frame interactive perception described in the present invention is also characterized in that the step 2.2 is performed according to the following steps:

[0031] Step 2.2.1: The initial feature F of the current frame t and the initial feature F of the history frame t-1 t-1 After pairing and splicing, the input is sent to the convolutional layer for processing to obtain the attention weight matrix T t-1 ;

[0032] Step 2.2.2: The attention weight matrix Tt-1 The size is reshaped to HW × 2, so that T t-1 Separate the data of two channels and use the data of the first channel as F t The attention weight T t-1,1 The data of the second channel is used as F t-1 The attention weight T t-1,2 ; Among them, H and W represent the initial features F t The height and width of

[0033] Step 2.2.3: Use formula (2) to get the initial feature F of the current frame t and the initial feature F of the history frame t-1 t-1 Aligned timing features ;

[0034] (2)

[0035] In formula (2), v t Indicates F t The vector to be learned, v t-1 Indicates F t-1 The vector to be learned; Represents the multiplication of the initial feature and the learning vector;

[0036] The step 3.3 is carried out as follows:

[0037] Step 3.3.1: Calculate according to formula (3) The corresponding query vector , thus calculating the key vector K and value vector V according to equation (4) and equation (5) respectively;

[0038] Q i = W d s i Q [ L N ( I i ) ] (3)

[0039] K = W d s K [ L N ( I ∑ ) ] (4)

[0040] V = W d s V [ L N ( I ∑ ) ] (5)

[0041] In formula (3)-(5), LN represents layer normalization; , and Represent the query vector , depthwise separable convolution operations corresponding to the key vector K and the value vector V; express Total embedded features after concatenation;

[0042] Step 3.3.2: Calculate according to formula (6) The corresponding attention weighted features ;

[0043] (6)

[0044] In formula (6), softmax represents the activation function; represents the impact factor, and ,in, Represents the total embedding feature The number of channels; I represents the instance normalization operation;

[0045] Step 3.3.3: and After adding, we get the i-th interaction layer feature .

[0046] The step 4 comprises:

[0047] Step 4.1: Use formula (7) to construct Weighted binary cross entropy loss ;

[0048] L W B C E , t = − 1 N ∑ n = 1 N ω n ⋅ [ g t , n ⋅ l o g p t , n + ( 1 − g t , n ) ⋅ ( 1 − l o g p t , n ) ] (7)

[0049] In formula (7), N represents The total number of pixels, and Respectively and p t The pixel value of the nth pixel in , Represents the weighted value of the nth pixel;

[0050] Step 4.2: Use formula (8) to construct Dice loss ;

[0051] (8)

[0052] Step 4.3: Use formula (9) to construct The maximum mean square error loss ;

[0053] (9)

[0054] In formula (9), It means to find the bi-norm of a matrix.

[0055] An electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the infrared weak small target detection method, and the processor is configured to execute the program stored in the memory.

[0056] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the infrared weak small target detection method when the computer program is run by a processor.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] 1. The present invention designs a simple and effective temporal feature extraction module for obtaining the temporal and spatial features of infrared weak small targets and improving the contrast of small targets. The module can effectively capture the motion information of the target and enhance the difference between the target and the background.

[0059] 2. This invention combines the advantages of Transformer and UNet, replacing the skip connection part of UNet with the multi-scale semantic interaction Transformer, and realizes cross-level information fusion. This improvement enables the network to better understand and utilize features at different levels.

[0060] 3. This paper proposes a multi-task weighted hybrid loss function that comprehensively considers classification accuracy, regional matching, and pixel-level prediction accuracy. This loss function can more comprehensively evaluate the performance of the model and guide the model optimization process, thereby improving the performance of the detection model in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a flow chart of the infrared small target detection method of the present invention;

[0062] Figure 2 It is a structural diagram of the time feature extraction module in the present invention;

[0063] Figure 3 This is the structural diagram of the feature interaction perception module in the present invention

[0064] Figure 4 This is a comparison diagram of the method of the present invention and other infrared target detection methods. DETAILED DESCRIPTION

[0065] In this embodiment, a method for detecting infrared dim small targets based on multi-frame interactive perception is designed by designing an effective temporal feature extraction module to obtain the temporal and spatial information of the target. In addition, the method designs a U-shaped interactive perception module and introduces a multi-scale semantic interactive Transformer block on a long-distance jump connection to achieve cross-level information fusion. At the same time, the method also proposes a multi-task weighted hybrid loss function to improve the performance of the model. Specifically, Figure 1 As shown, the steps are as follows:

[0066] Step 1: preprocessing the infrared sequence images;

[0067] After obtaining the infrared image sequence and performing preprocessing such as resizing, cropping, horizontal and vertical flipping, the processed infrared image sequence is obtained, which is denoted as ;in, Represents the infrared image of the historical frame tn. When n=0, represents the infrared image of the current frame; k represents the number of frames in the infrared sequence image; let The true label is recorded as ; In this embodiment, the size of each image is cropped to 256×256.

[0068] Step 2: Build a temporal feature extraction module for continuous multi-frame images, the structure is as follows Figure 2 As shown, it includes: center difference convolution block, time attention module and residual convolution block;

[0069] Step 2.1: The central difference convolution block uses formula (1) to The purpose of the calculation is to extract the center-oriented gradient perception features and enhance the contrast of small targets to obtain the initial feature F of the infrared image of the tnth frame. t-n , thus obtaining the initial features of the infrared image sequence ; Among them, when n=0, F t Indicates the infrared image of the current frame The initial characteristics of

[0070] (1)

[0071] In formula (1), is an integer between [-1,1], Respectively represent the offset of the horizontal and vertical coordinates; express The pixel value at the center position (x, y) of express Middle position The pixel value at It's location The weights to be learned at is a hyperparameter, and θ ∈ [ 0 , 1 ] In this embodiment, Set to 0.7.

[0072] Step 2.2: Temporal attention module Processing, the initial feature F of the current frame t and the initial feature F of the history frame t-1 t-1After pairing, the temporal features of the historical frame t-1 are obtained , thus obtaining the temporal characteristics of the infrared image sequence ;

[0073] Step 2.2.1: The initial feature F of the current frame t and the initial feature F of the history frame t-1 t-1 After pairing and splicing, the input is sent to the convolutional layer for processing to obtain the attention weight matrix T t-1 ;

[0074] Step 2.2.2: The attention weight matrix T t-1 The size is reshaped to HW × 2, so that T t-1 Separate the data of two channels and use the data of the first channel as F t The attention weight T t-1,1 The data of the second channel is used as F t-1 The attention weight T t-1,2 ; Among them, H and W represent the initial features F t The height and width of the

[0075] Step 2.2.3: Use formula (2) to get the initial feature F of the current frame t and the initial feature F of the history frame t-1 t-1 Aligned timing features ;

[0076] (2)

[0077] In formula (2), v t Indicates F t The vector to be learned, v t-1 Indicates F t-1 The vector to be learned; Indicates the multiplication of the initial feature and the learning vector; in this embodiment, the current frame image and each historical frame image are processed through the center difference convolution and time attention module, and the weights are shared, so that the common significant target features can be obtained. -

[0078] Step 2.3: The temporal features of the infrared image sequence and the initial feature F of the current frame t After splicing, it is input into the residual convolution block for processing to obtain the infrared image of the current frame Temporal and spatial characteristics of small mid-infrared targets st ;

[0079] Step 3: Build a U-shaped feature interaction perception module with the following structure: Figure 3As shown in Figure 1, it includes: feature interaction perception module encoder, decoder and multi-scale semantic interaction Transformer block, and is used to st After processing, the interactive perception feature F of infrared weak targets is obtained A ;

[0080] Step 3.1: The feature interaction perception module encoder uses m groups of residual convolution blocks containing downsampling to generate F st Processing to obtain high-level features of the target ,in, represents the target high-level features output by the i-th residual convolution block, C i is the channel dimension of the output of the i-th group of residual convolution blocks; In this embodiment, m=3, = 32, = 64, = 128.

[0081] Step 3.2: Use convolutions of different sizes to verify the high-level features of the target Perform patch embedding to obtain a series of embedding layer features ,in, represents the i-th embedding layer feature, and h represents The width of , w represents The height of ; In this embodiment, the sizes of the convolution kernels are 16, 8, and 4 respectively.

[0082] Step 3.3: Use the multi-scale semantic interaction Transformer block to embed the features of the layer Processing to obtain the interactive layer features ,in, represents the i-th interaction layer feature;

[0083] Step 3.3.1: Calculate according to formula (3) The corresponding query vector , thus calculating the key vector K and value vector V according to equation (4) and equation (5) respectively;

[0084] Q i = W d s i Q [ L N ( I i ) ] (3)

[0085] K = W d s K [ L N ( I ∑ ) ] (4)

[0086] V = W d s V [ L N ( I ∑ ) ] (5)

[0087] In formula (3)-(5), LN represents layer normalization; , and Represent the query vector , depthwise separable convolution operations corresponding to the key vector K and the value vector V; express Total embedded features after concatenation.

[0088] Step 3.3.2: Calculate according to formula (6) The corresponding attention weighted features ;

[0089] (6)

[0090] In formula (6), softmax represents the activation function; represents the impact factor, and ,in, Represents the total embedding feature The number of channels; I represents the instance normalization operation;

[0091] Step 3.3.3: and After adding, we get the i-th interaction layer feature .

[0092] Step 3.4: Use the decoder to decode the interactive layer feature sequence Perform decoding operation to obtain the interactive perception feature F of infrared weak target A ; In this embodiment, the feature tensor after multi-scale semantic interaction is input into the corresponding decoder, which is similar to the upsampling decoding process on the U-Net network.

[0093] Step 3.5: F A After being processed by residual convolution blocks and 1×1 convolutions, we get The infrared prediction results of t .

[0094] Step 4: Establish a hybrid loss function;

[0095] Step 4.1: Use formula (7) to construct Weighted binary cross entropy loss ;

[0096] L W B C E , t = − 1 N ∑ n = 1 N ω n ⋅ [ g t , n ⋅ l o g p t , n + ( 1 − g t , n ) ⋅ ( 1 − l o g p t , n ) ] (7)

[0097] In formula (7), N represents The total number of pixels, and Respectively and p t Middle The pixel value of pixels, Represents the weighted value of the nth pixel.

[0098] Step 4.2: Use formula (8) to construct Dice loss ;

[0099] (8)

[0100] Step 4.3: Use formula (9) to construct The maximum mean square error loss ;

[0101] (9)

[0102] In formula (9), It means to find the bi-norm of a matrix.

[0103] Step 4.4: Use formula (10) to construct The total loss function L t ;

[0104] (10)

[0105] In formula (10), express The weighted binary cross entropy loss of ; express The maximum mean square error loss of express Dice loss; , are two weighting coefficients; in this embodiment, the weighting coefficients are =10, =0.2;

[0106] Step 5: Based on , the network composed of the time series feature extraction module and the feature interaction perception module is trained using the gradient descent algorithm, and the total loss function L is calculated to update the network parameters. When the number of training iterations reaches the set number, the training is stopped to obtain the trained infrared weak small target detection model, which is used to detect infrared weak small targets under complex backgrounds. In this embodiment, the Adam optimizer is used to randomly extract 175 groups of sequences from the SIATD complex background infrared weak moving target data set for training, and 175 groups for testing. The training is stopped after 200 iterations.

[0107] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0108] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

[0109] Figure 4 The qualitative comparison results between the present invention and other methods in complex backgrounds are shown. Figure 4 The first column shows the original infrared image, the second to sixth columns show the detection results of five single-frame deep learning methods, including asymmetric context modulation method (ACM), attention local contrast network (ALCNet), residual connection based UNet network (Res-UNet), dense nested attention network (DNANet) and UNet nested network (UIUNet), the seventh column shows the detection results of the method proposed in this invention, and the last column shows the true label of the original infrared image. Among them, the red circle represents the correctly detected target, the blue circle represents the missed target, and the yellow circle represents the misdetected target. It can be seen that most single-frame based deep learning methods will produce many false alarms and missed detections when facing complex backgrounds, and even if the target is detected, its outline is usually blurred, which hinders further accurate recognition of the target type. The present invention constructs a temporal feature extraction module for continuous multi-frame images and a feature interaction perception module with a U-shaped structure. The results show that this method can not only detect faint targets using spatial information, but also identify targets that are difficult to detect based on spatial information alone through temporal series clues. Not only local contrast information is constructed, but also the ability of long-range image dependencies is established, making it possible to completely and accurately segment the shape of the target.

Claims

1. A method for detecting infrared small targets based on multi-frame interactive perception, characterized in that: The steps are as follows: Step 1: preprocessing the infrared sequence images; After obtaining the infrared image sequence and performing preprocessing such as resizing, cropping, horizontal and vertical flipping, the processed infrared image sequence is obtained, which is denoted as ;in, Represents the infrared image of the tnth frame in history. When n=0, let represents the infrared image of the current frame; k represents the number of frames in the infrared sequence image; let The true label is recorded as ; Step 2: Construct a temporal feature extraction module for continuous multi-frame images, including: center difference convolution block, temporal attention module and residual convolution block; Step 2.1: The central difference convolution block uses formula (1) to Calculate and get the initial feature F of the infrared image of the tnth frame t-n , thus obtaining the initial features of the infrared image sequence ; Among them, when n=0, F t Indicates the infrared image of the current frame The initial characteristics of (1) In formula (1), is an integer between [-1,1], Respectively represent the offset of the horizontal and vertical coordinates; express The pixel value at the center position (x, y) of express Middle position The pixel value at It's location The weights to be learned at is a hyperparameter, and ; Step 2.2: Temporal attention module Processing, the initial feature F of the current frame t and the initial feature F of the history frame t-1 t-1 After pairing, the temporal features of the historical frame t-1 are obtained , thus obtaining the temporal characteristics of the infrared image sequence ; Step 2.3: The temporal features of the infrared image sequence and the initial feature F of the current frame t After splicing, it is input into the residual convolution block for processing to obtain the infrared image of the current frame Temporal and spatial characteristics of small mid-infrared targets st ; Step 3: Construct a U-shaped feature interaction perception module, including: feature interaction perception module encoder, decoder and multi-scale semantic interaction Transformer block, and use it to st After processing, the interactive perception feature F of infrared weak targets is obtained A ; Step 3.1: The feature interaction perception module encoder uses m groups of residual convolution blocks containing downsampling to generate F st Processing to obtain high-level features of the target ,in, represents the target high-level features output by the i-th residual convolution block, C i is the channel dimension of the output of the i-th group of residual convolution blocks; Step 3.2: Use convolutions of different sizes to verify the high-level features of the target Perform patch embedding to obtain a series of embedding layer features ,in, represents the i-th embedding layer feature, and h represents The width of , w represents The height of ; Step 3.3: Use the multi-scale semantic interaction Transformer block to embed the features of the layer Processing to obtain the interactive layer features ,in, represents the i-th interaction layer feature; Step 3.4: Use the decoder to decode the interactive layer feature sequence Perform decoding operation to obtain the interactive perception feature F of infrared weak target A ; Step 3.5: F A After processing by residual convolution block and 1×1 convolution in sequence, we get The infrared prediction results p t ; Step 4: Use formula (10) to establish The total loss function L t ; (10) In formula (10), express The weighted binary cross entropy loss of ; express The maximum mean square error loss; express Dice loss; , are two weighting coefficients; Step 5: Based on The network composed of the time series feature extraction module and the feature interaction perception module is trained using the gradient descent algorithm, and the total loss function L is calculated to update the network parameters. When the number of training iterations reaches the set number, the training is stopped to obtain the trained infrared dim small target detection model, which is used to realize the detection of infrared dim small targets in complex backgrounds.

2. The infrared small target detection method based on multi-frame interactive perception according to claim 1 is characterized in that: The step 2.2 is performed as follows: Step 2.2.1: The initial feature F of the current frame t and the initial feature F of the history frame t-1 t-1 After pairing and splicing, the input is sent to the convolutional layer for processing to obtain the attention weight matrix T t-1 ; Step 2.2.2: The attention weight matrix T t-1 The size is reshaped to HW × 2, so that T t-1 Separate the data of two channels and use the data of the first channel as F t The attention weight T t-1,1 The data of the second channel is used as F t-1 The attention weight T t-1,2 ; Among them, H and W represent the initial features F t The height and width of Step 2.2.3: Use formula (2) to get the initial feature F of the current frame t and the initial feature F of the history frame t-1 t-1 Aligned timing features ; (2) In formula (2), v t Indicates F t The vector to be learned, v t-1 Indicates F t-1 The vector to be learned; Represents the multiplication of the initial features and the learning vector.

3. The infrared small target detection method based on multi-frame interactive perception according to claim 1 is characterized in that: The step 3.3 is carried out as follows: Step 3.3.1: Calculate according to formula (3) The corresponding query vector , thus calculating the key vector K and value vector V according to equation (4) and equation (5) respectively; (3) (4) (5) In formula (3)-(5), LN represents layer normalization; , and Represent the query vector , depthwise separable convolution operations corresponding to the key vector K and the value vector V; express Total embedded features after concatenation; Step 3.3.2: Calculate according to formula (6) The corresponding attention weighted features ; (6) In formula (6), softmax represents the activation function; represents the impact factor, and ,in, Represents the total embedding feature The number of channels; I represents the instance normalization operation; Step 3.3.3: and After adding, we get the i-th interaction layer feature .

4. The infrared small target detection method based on multi-frame interactive perception according to claim 3 is characterized in that: The step 4 comprises: Step 4.1: Use formula (7) to construct Weighted binary cross entropy loss ; (7) In formula (7), N represents The total number of pixels, and Respectively and p t The pixel value of the nth pixel in , Represents the weighted value of the nth pixel; Step 4.2: Use formula (8) to construct Dice loss ; (8) Step 4.3: Use formula (9) to construct The maximum mean square error loss ; (9) In formula (9), It means to find the bi-norm of a matrix.

5. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute any one of the infrared dim small target detection methods in claims 1-4, and the processor is configured to execute the program stored in the memory.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the infrared dim small target detection method according to any one of claims 1 to 4 are executed.

Images