An infrared dim small target detection method based on multi-frame interactive perception

Abstract

The application discloses an infrared dim small target detection method based on multi-frame interactive perception, comprising the following steps: 1, pre-processing the infrared sequence image to avoid overfitting in the training process; 2, constructing a time sequence feature extraction module for continuous multi-frame images; 3, constructing a feature interactive perception module with a U-shaped structure; 4, establishing a loss function between the result of the current frame of infrared image after the prediction module and the real binary image; 5, training the network with the infrared image, and detecting the infrared dim small target with the trained model. The application utilizes multi-frame images to obtain the space-time information of the target, and realizes cross-level information fusion, thereby improving the accuracy of the infrared dim small target detection, and particularly enhancing the capture ability of the fast-moving small target.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of computer vision, and particularly relates to an infrared dim small target detection method based on multi-frame interactive perception. BACKGROUND

[0002] Infrared imaging is widely used in civil, military, industrial and remote sensing fields due to its strong anti-interference ability to light changes. Infrared small target (IRST) detection algorithm, as a key technology in this field, has attracted attention from scholars at home and abroad. Its purpose is to accurately locate small targets in various infrared backgrounds. However, due to the influence of target radiation characteristics and transmission path, small targets in infrared images often exhibit 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 multi-filter, sparse low-rank and human visual system (HVS) methods, although have achieved results in specific fields, are highly dependent on manual feature design and prior knowledge. When facing dramatic changes in scenes, these traditional methods are difficult to adapt to new environments due to fixed hyperparameters, and their performance is significantly reduced. In contrast, deep learning-based methods exhibit strong modeling capabilities, extracting highly discriminative features through learning, and significantly improving performance. 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 single-frame detectors to process multiple frames of input, trying to retain dynamic information in the converted single-frame image. However, these methods are not designed for multi-frame scenes, so there is still room for exploration in the effective use of temporal information. In addition, with the rise of Vision Transformer (ViT), which decomposes images into patches and calculates correlations, it is widely used in global image modeling.

[0005] Although significant progress has been made in deep learning-based target detection methods, existing infrared small target detection methods still have the following shortcomings:

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

[0007] 2. Existing multi-frame IRST detection methods are not designed specifically for multi-frame scenes, so there is still room for exploration in the use of temporal information.

[0008] 3.Compared with the improved network structure, the current method only builds long-distance dependency on a single feature map in the application of spatial self-attention (SA), rather than establishing context across layers. SUMMARY

[0009] The present application is to solve the above-mentioned deficiencies in the 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, so as to suppress background noise, retain the edge contour features of the infrared dim small target, and enhance the capture ability of the fast-moving small target.

[0010] In order to achieve the above-mentioned application purposes, the present application adopts the following technical solutions:

[0011] The infrared dim small target detection method based on multi-frame interactive perception has the characteristics that the following steps are performed:

[0012] Step 1, pre-processing the infrared sequence image;

[0013] After obtaining the infrared image sequence and performing size unification, cropping, horizontal and vertical flip preprocessing, the processed infrared image sequence is obtained, denoted as ; wherein represents the historical t-n frame infrared image, when n=0, let represents the current frame infrared image; k represents the frame number in the infrared sequence image; let The true label of is denoted as ;

[0014] Step 2, constructing a time sequence feature extraction module for continuous multiple frames of images, including: a central difference convolution block, a time attention module and a residual convolution block;

[0015] Step 2.1, the central difference convolution block calculates the t-n frame infrared image using formula (1) on to obtain the initial feature F t-n of the t-n frame infrared image, so as to obtain the initial feature of the infrared image sequence ; wherein, when n=0, F t represents the initial feature of the current frame infrared image ;

[0016] (1)

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

[0018] Step 2.2, Time Attention Module Processing is performed to obtain the initial features F of the current frame. t The initial feature F of the historical (t-1)th frame t-1 After pairing, the temporal features of the historical frame t-1 are obtained. Thus, the temporal characteristics of the infrared image sequence are obtained. ;

[0019] Step 2.3: Extract the temporal features of the infrared image sequence. With the initial feature F of the current frame t After stitching, the images are input into a residual convolutional block for processing to obtain the infrared image of the current frame. Spatiotemporal characteristics F of weak mid-infrared targets st ;

[0020] Step 3: Construct a U-shaped feature interaction perception module, including: a feature interaction perception module encoder, decoder, and a multi-scale semantic interaction Transformer block, and use it to process F... st The data is processed to obtain the interactive sensing features F of the infrared weak target. A ;

[0021] Step 3.1: The feature interaction perception module encoder uses m sets of residual convolutional blocks containing downsampling to pair F. st Processing is performed to obtain the high-level features of the target. ,in, C represents the target high-level feature output by the i-th residual convolutional block. i Let be the channel dimension of the output of the i-th residual convolutional block;

[0022] Step 3.2: Use convolutional kernels of different sizes to extract high-level features of the target. Patch embedding is performed to obtain a series of embedding layer features. ,in, Let h represent the feature of the i-th embedding layer. The width, and w represents The height, and ;

[0023] Step 3.3: Use multi-scale semantic interaction Transformer blocks to apply features from the embedding layer. Processing is performed to obtain interaction layer features. ,in, Represents the feature of the i-th interaction layer;

[0024] Step 3.4: Use the decoder to process the interaction layer feature sequence. Decoding is performed to obtain the interactive sensing features F of the infrared weak target. A ;

[0025] Step 3.5, F A After processing through residual convolution blocks and 1×1 convolutions in sequence, we obtain... Infrared prediction results p t ;

[0026] Step 4: Establish using equation (10) Total loss function L t ;

[0027] (10)

[0028] In equation (10), express The weighted binary cross-entropy loss; express The maximum mean squared error loss; express Dice loss; , There are two weighting coefficients;

[0029] Step 5, based on The gradient descent algorithm is used to train the network consisting of the temporal feature extraction module and the feature interaction perception module, 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, thus obtaining the trained infrared weak target detection model, which is used to detect infrared weak targets in complex backgrounds.

[0030] The infrared weak target detection method based on multi-frame interactive sensing described in this invention is characterized in that step 2.2 is performed as follows:

[0031] Step 2.2.1: Set the initial features F of the current frame t The initial feature F of the historical (t-1)th frame t-1 After pairing and concatenation, the input is fed into a convolutional layer for processing to obtain the attention weight matrix T. t-1 ;

[0032] Step 2.2.2: Set the attention weight matrix Tt-1 is reshaped to HWx2, so that from T t-1 , the data of two channels are separated, and the data of the first channel is used as the attention weight T t of F t-1,1 , and the data of the second channel is used as the attention weight T t-1 of F t-1,2 ; wherein H and W respectively represent the height and width of the initial feature F t ;

[0033] Step 2.2.3, using formula (2) to obtain the initial feature F t of the current frame aligned with the initial feature F t-1 of the historical t-1 frame ;

[0034] (2)

[0035] In formula (2), v t represents the to-be-learned vector of F t , and v t-1 represents the to-be-learned vector of F t-1 ; represents that the initial feature is multiplied by the learning vector;

[0036] The step 3.3 is performed as follows:

[0037] Step 3.3.1, according to formula (3), calculate the corresponding query vector , so as to calculate the key vector K and the value vector V according to formula (4) and formula (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 respectively represent the depth separable convolution operation corresponding to the query vector , the key vector K and the value vector V; represents the total embedding feature after splicing;

[0042] Step 3.3.2, according to formula (6), calculate Corresponding attention weighted features ;

[0043] (6)

[0044] In formula (6), softmax represents an activation function; represents an influence factor, and wherein, represents a total embedding feature ; I represents an instance normalization operation;

[0045] Step 3.3.3, after adding and , the i th interaction layer feature is obtained.

[0046] The step 4 comprises:

[0047] Step 4.1, a weighted binary cross-entropy loss of formula (7) is constructed ;

[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 a total pixel number of , and respectively represent pixel values of the n th pixel in and p t , and represents a weighted value of the n th pixel.

[0050] Step 4.2, a Dice loss of formula (8) is constructed ;

[0051] (8)

[0052] Step 4.3, a maximum mean square error loss of formula (9) is constructed ;

[0053] (9)

[0054] In formula (9), represents a two-norm of a matrix.

[0055] The electronic device comprises a memory and a processor, and is characterized in that the memory is used for storing a program supporting the processor to execute the infrared dim target detection method, and the processor is configured to execute the program stored in the memory.

[0056] The computer readable storage medium of the present application, the computer readable storage medium stores a computer program, characterized in that, the computer program is executed by the processor to perform the steps of the infrared dim target detection method.

[0057] Compared with the prior art, the present application has the advantages that:

[0058] 1、The present application designs a simple and effective timing feature extraction module for obtaining the space-time features of infrared dim targets and improving the contrast of small targets. This module can effectively capture the motion information of the target and enhance the difference between the target and the background.

[0059] 2、The present application combines the advantages of Transformer and UNet, replaces the skip connection part of UNet with a multi-scale semantic interaction Transformer to realize cross-level information fusion. This improvement enables the network to better understand and utilize features at different levels.

[0060] 3、The present application proposes a multi-task weighted hybrid loss function that considers classification accuracy, region matching degree, 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 scenes. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 The infrared dim target detection method flowchart in the present application;

[0062] Figure 2 The time feature extraction module structure diagram in the present application;

[0063] Figure 3 The feature interaction perception module structure diagram in the present application

[0064] Figure 4 The comparison results of the present application method and other infrared target detection methods. DETAILED DESCRIPTION

[0065] In this embodiment, an infrared dim target detection method based on multi-frame interaction perception is designed to extract the space-time information of the target through an effective timing feature extraction module. In addition, the method designs an interactive perception module with a U-shaped structure, which introduces a multi-scale semantic interaction Transformer block in the long-distance jump connection to realize 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, as shown in Figure 1 the following steps are performed:

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

[0067] After acquiring the infrared image sequence and performing preprocessing such as resizing, cropping, and horizontal and vertical flipping, the processed infrared image sequence is obtained, denoted as... ;in, This represents the infrared image of the nth frame in history. When n=0, This represents the infrared image of the current frame; k represents the frame number in the infrared image sequence; let... The real label is recorded as In this embodiment, each image is cropped to a size of 256×256.

[0068] Step 2: Construct a temporal feature extraction module for consecutive multi-frame images, with the following structure: Figure 2 As shown, it includes: a central difference convolutional block, a temporal attention module, and a residual convolutional block;

[0069] Step 2.1, using equation (1) to perform central difference convolution blocks. The calculations are performed to extract center-oriented gradient sensing features and enhance the contrast of small targets, resulting in the initial feature F of the tn-th frame of the infrared image. t-n This allows us to obtain the initial features of the infrared image sequence. ; where, when n=0, F t Indicates the infrared image of the current frame. Initial characteristics;

[0070] (1)

[0071] In equation (1), It is an integer belonging to the range [-1, 1]. These represent the offsets of the horizontal and vertical axes, respectively. express The pixel value at the center position (x, y). express Middle position Pixel value at that location, It is a location The weights to be learned at the location, It is a hyperparameter, and θ ∈ [ 0 , 1 ] In this embodiment, Set it to 0.7.

[0072] Step 2.2, Time Attention Module Processing is performed to obtain the initial features F of the current frame. t The initial feature F of the historical (t-1)th frame t-1After the pairing processing, the time sequence feature of the historical t-1 frame is obtained , thereby obtaining the time sequence feature of the infrared image sequence ;

[0073] Step 2.2.1, the initial feature F t of the current frame is matched and spliced with the initial feature F t-1 of the historical t-1 frame t-1 ;

[0074] Step 2.2.2, the size of the attention weight matrix T t-1 is reshaped to H*W*2, thereby separating the data of two channels from T t-1 , and using the data of the first channel as the attention weight T t of F t-1,1 , and the data of the second channel as the attention weight T t-1 of F t-1,2 ; wherein H and W represent the height and width of the initial feature F t , respectively.

[0075] Step 2.2.3, the time sequence feature F t of the initial feature F t-1 of the current frame and the initial feature F of the historical t-1 frame after alignment is obtained using formula (2)

[0076] (2)

[0077] In formula (2), v t represents the to-be-learned vector of F t , and v t-1 represents the to-be-learned vector of F t-1 ; represents the multiplication of the initial feature and the learned vector; in this embodiment, the current frame image and each historical frame image are processed by the center difference convolution and the time attention module, share the weights, and thus the common salient target feature can be obtained.

[0078] Step 2.3, after the time sequence feature F of the infrared image sequence and the initial feature F t of the current frame are spliced, the spliced feature is input into the residual convolution block for processing, and the space-time feature F st of the infrared weak small target in the infrared image of the current frame is obtained ;

[0079] Step 3, a feature interaction perception module with a U-shaped structure is constructed, and the structure is as shown in Figure 3As shown, it includes: a feature interaction perception module encoder, decoder, and a multi-scale semantic interaction Transformer block, and is used for F... st The data is processed to obtain the interactive sensing features F of the infrared weak target. A ;

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

[0081] Step 3.2: Use convolutional kernels of different sizes to extract high-level features of the target. Patch embedding is performed to obtain a series of embedding layer features. ,in, Let h represent the feature of the i-th embedding layer. The width, and w represents The height, and In this embodiment, the sizes of the convolution kernels are 16, 8, and 4, respectively.

[0082] Step 3.3: Use multi-scale semantic interaction Transformer blocks to apply features from the embedding layer. Processing is performed to obtain interaction layer features. ,in, Represents the feature of the i-th interaction layer;

[0083] Step 3.3.1: Calculate according to formula (3) Corresponding query vector Thus, the key vector K and the value vector V are calculated according to equations (4) and (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 equations (3)-(5), LN represents layer normalization; , and These represent the query vectors respectively. Depth-separable convolution operations correspond to the key vector K and the value vector V; express The total embedded features after concatenation.

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

[0089] (6)

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

[0091] Step 3.3.3, will and After addition, the i-th interaction layer feature is obtained. .

[0092] Step 3.4: Use the decoder to process the interaction layer feature sequence. Decoding is performed to obtain the interactive sensing features F of the 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 sampling and decoding process on the U-Net network.

[0093] Step 3.5, F A After processing through residual convolution blocks and 1×1 convolutions in sequence, we obtain... Infrared prediction results p t .

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

[0095] Step 4.1: Construct using equation (7) 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 equation (7), N represents Total number of pixels, and They represent and p t The Middle The pixel value of each pixel. This represents the weighted value of the nth pixel.

[0098] Step 4.2: Construct using equation (8) Dice loss ;

[0099] (8)

[0100] Step 4.3: Construct using equation (9) Maximum mean squared error loss ;

[0101] (9)

[0102] In equation (9), This indicates the calculation of the 2-norm of a matrix.

[0103] Step 4.4: Construct using equation (10) Total loss function L t ;

[0104] (10)

[0105] In equation (10), express The weighted binary cross-entropy loss; express The maximum mean squared error loss; express Dice loss; , There are two weighting coefficients; in this embodiment, the weighting coefficients are respectively =10, =0.2;

[0106] Step 5, based on The gradient descent algorithm is used to train the network consisting of the temporal feature extraction module and the feature interaction perception module, and the total loss function L is calculated to update the network parameters. Training is stopped when the number of training iterations reaches a set number, thus obtaining the trained infrared weak target detection model, which is used to detect infrared weak targets in complex backgrounds. In this embodiment, the Adam optimizer is used, and 175 sets of sequences are randomly extracted from the SIATD infrared weak moving target dataset in complex backgrounds for training and 175 sets are used for testing. Training is stopped after 200 iterations.

[0107] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described 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, and the computer program is run by a processor to execute the steps of the above method.

[0109] Figure 4 Qualitative comparison results of the present application and other methods in complex background are shown. Figure 4 The first column in the middle shows the original infrared image, and 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), UNet network based on residual connection (Res-UNet), dense nested attention network (DNANet) and UNet nested network (UIUNet), the seventh column shows the detection result of the method proposed in the present application, 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 false detection target. It can be seen that most of the single-frame-based deep learning methods will produce many false positives and missed detections when facing complex backgrounds, and even if the target is detected, its outline is usually fuzzy, which hinders further accurate identification of the target type. The present application constructs a time sequence feature extraction module for continuous multiple frames of images and a U-shaped structure feature interaction perception module. The results show that the present method not only can detect weak targets by using spatial information, but also can identify those targets that are difficult to detect by spatial information alone through time sequence clues. Not only the local contrast information is constructed, but also the ability to establish image long-range dependency makes it possible to completely and accurately segment the shape of the target.

Claims

1. A method for detecting weak infrared targets based on multi-frame interactive sensing, characterized in that, The procedure is as follows: Step 1: Preprocess the infrared sequence images; After acquiring the infrared image sequence and performing preprocessing such as resizing, cropping, and horizontal and vertical flipping, the processed infrared image sequence is obtained, denoted as... ;in, This represents the infrared image of the nth frame in history. When n=0, let... This represents the infrared image of the current frame; k represents the frame number in the infrared image sequence; let... The real label is recorded as ; Step 2: Construct a temporal feature extraction module for consecutive multi-frame images, including: a center difference convolutional block, a temporal attention module, and a residual convolutional block; Step 2.1: The central difference convolution block is processed using equation (1). Calculations are performed to obtain the initial features F of the tn-th frame of the infrared image. t-n This allows us to obtain the initial features of the infrared image sequence. ; where, when n=0, F t Indicates the infrared image of the current frame. Initial characteristics; (1) In equation (1), It is an integer belonging to the range [-1, 1]. These represent the offsets of the horizontal and vertical axes, respectively. express The pixel value at the center position (x, y). express Middle position Pixel value at that location, It is a location The weights to be learned at the location, It is a hyperparameter, and ; Step 2.2, Time Attention Module Processing is performed to obtain the initial features F of the current frame. t The initial feature F of the historical (t-1)th frame t-1 After pairing, the temporal features of the historical frame t-1 are obtained. Thus, the temporal characteristics of the infrared image sequence are obtained. ; Step 2.3: Extract the temporal features of the infrared image sequence. With the initial feature F of the current frame t After stitching, the images are input into a residual convolutional block for processing to obtain the infrared image of the current frame. Spatiotemporal characteristics F of weak mid-infrared targets st ; Step 3: Construct a U-shaped feature interaction perception module, including: a feature interaction perception module encoder, decoder, and a multi-scale semantic interaction Transformer block, and use it to process F... st The data is processed to obtain the interactive sensing features F of the infrared weak target. A ; Step 3.1: The feature interaction perception module encoder uses m sets of residual convolutional blocks containing downsampling to pair F. st Processing is performed to obtain the high-level features of the target. ,in, C represents the target high-level feature output by the i-th residual convolutional block. i Let be the channel dimension of the output of the i-th residual convolutional block; Step 3.2: Use convolutional kernels of different sizes to extract high-level features of the target. Patch embedding is performed to obtain a series of embedding layer features. ,in, Let h represent the feature of the i-th embedding layer. The width, and w represents The height, and ; Step 3.3: Use multi-scale semantic interaction Transformer blocks to apply features from the embedding layer. Processing is performed to obtain interaction layer features. ,in, Represents the feature of the i-th interaction layer; Step 3.4: Use the decoder to process the interaction layer feature sequence. Decoding is performed to obtain the interactive sensing features F of the infrared weak target. A ; Step 3.5, F A After processing through residual convolution blocks and 1×1 convolutions in sequence, we obtain... Infrared prediction results p t ; Step 4: Establish using equation (10) Total loss function L t ; (10) In equation (10), express The weighted binary cross-entropy loss; express The maximum mean squared error loss; express Dice loss; , There are two weighting coefficients; Step 5, based on The gradient descent algorithm is used to train the network consisting of the temporal feature extraction module and the feature interaction perception module, 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, thus obtaining the trained infrared weak target detection model, which is used to detect infrared weak targets in complex backgrounds.

2. The infrared weak target detection method based on multi-frame interactive sensing according to claim 1, characterized in that, Step 2.2 is performed as follows: Step 2.2.1: Set the initial features F of the current frame t The initial feature F of the historical (t-1)th frame t-1 After pairing and concatenation, the input is fed into a convolutional layer for processing to obtain the attention weight matrix T. t-1 ; Step 2.2.2: Set the attention weight matrix T t-1 The dimensions are reshaped to HW×2, thus from T t-1 The data from the first channel is separated into two channels, and the data from the first channel is used as F. t Attention weight T t-1,1 The data from the second channel is used as F t-1 Attention weight T t-1,2 Where H and W represent the initial features F, respectively. t Height and width; Step 2.2.3: Use equation (2) to obtain the initial feature F of the current frame. t The initial feature F of the historical (t-1)th frame t-1 Aligned temporal features ; (2) In equation (2), v t F represents t The vector to be learned, v t-1 F represents t-1 The vector to be learned; This indicates that the initial features are multiplied by the learning vector.

3. The infrared weak target detection method based on multi-frame interactive sensing according to claim 1, characterized in that, Step 3.3 is performed as follows: Step 3.3.1: Calculate according to formula (3) Corresponding query vector Thus, the key vector K and the value vector V are calculated according to equations (4) and (5) respectively; (3) (4) (5) In equations (3)-(5), LN represents layer normalization; , and These represent the query vectors respectively. Depth-separable convolution operations correspond to the key vector K and the value vector V; express The concatenated total embedding features; Step 3.3.2: Calculate according to formula (6) Corresponding attention-weighted features ; (6) In equation (6), softmax represents the activation function; Indicates the impact factor, and ,in, Represents total embedding features The number of channels; I represents the instance normalization operation; Step 3.3.3, will and After addition, the i-th interaction layer feature is obtained. .

4. The infrared weak target detection method based on multi-frame interactive sensing according to claim 3, characterized in that, Step 4 includes: Step 4.1: Construct using equation (7) Weighted binary cross-entropy loss ; (7) In equation (7), N represents Total number of pixels, and They represent and p t The pixel value of the nth pixel. This represents the weighted value of the nth pixel; Step 4.2: Construct using equation (8) Dice loss ; (8) Step 4.3: Construct using equation (9) Maximum mean squared error loss ; (9) In equation (9), This indicates the calculation of the 2-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 in executing any of the infrared weak target detection methods of claims 1-4, and the processor is configured to execute the program stored in the memory.

6. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is run by the processor, it performs the steps of any one of the infrared weak target detection methods described in claims 1-4.

Images