Small target detection method based on curvature attention mechanism

Abstract

The present invention proposes a small target detection method based on a curvature attention mechanism, and the implementation steps are as follows: obtaining a training sample set and a test sample set; constructing a small target detection network model based on the curvature attention mechanism; initializing parameters; training the target detection network model; updating the parameters of the target detection network model; and obtaining the trained target detection network model and the detection results of the small target. The backbone network of the present invention can extract small target features of various scales and reduce feature loss. The attention map of each sample obtained by the curvature network can supplement the fuzzy edge prediction results in the segmentation map. The attention network fuses the features of each segmentation map obtained by the backbone network with each attention map obtained by the corresponding curvature network, avoiding the defects of the existing technology network in paying insufficient attention to key edge features and incomplete extraction of target information. The test results show that the present invention can effectively improve the accuracy of small target detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of digital image processing and relates to a method for detecting small targets in an image, and specifically to a small target detection method based on a curvature attention mechanism, which can be used for maritime rescue and forest fire fighting. Background Art

[0002] Small target detection uses computer technology to detect objects from blurred images. Small target detection has a wide range of applications, such as forest firefighting and maritime rescue. Compared to general target detection scenarios, small targets have the following characteristics: In different scenarios, the target size varies from a few pixels to hundreds of pixels due to the large variations in shooting distance. The targets are diverse, often objects with distinct shapes, such as aircraft and ships. These objects vary significantly in shape at different angles and distances, resulting in highly irregular shapes in the image. The image background is often subject to significant noise and clutter, which can easily hide the target, resulting in low contrast and signal-to-noise ratio. To detect small targets, researchers have proposed several methods based on image processing and machine learning, including those based on filtering, the human visual system, and low-rank sparse matrices. However, these traditional methods have several limitations: they cannot suppress complex background noise, resulting in high detection error rates and unstable performance.

[0003] With the success of deep learning in many fields, researchers have applied it to small target detection, achieving significant results. For example, Beijing Institute of Technology proposed a context-based infrared small target detection method in its patent application, "A Method for Detecting Infrared Small Targets Based on Contextual Information" (patent application number: 202211461433.8, publication number: CN115797684A). This method first uses a convolutional neural network to extract image features, then performs feature fusion, and finally detects the target based on the fused features. The feature extraction stage considers the contextual information requirements of targets of varying shapes, dynamically capturing contextual information to establish long-range dependencies between features. To optimize features, the method divides the input into sequences, incorporates positional information, and feeds it into a multi-head attention mechanism. To further enhance feature representation, the method uses residual connections and convolutional layers. The feature fusion stage aggregates channel and spatial information from multiple source features to enhance subsequent detection. Methodologically, the algorithm has certain innovations, such as dynamic contextual feature extraction and multi-head attention feature optimization. However, the network's insufficient focus on key edge features and incomplete extraction of target information results in limited detection performance. Summary of the Invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a small target detection method based on the curvature attention mechanism, which is used to solve the technical problem of low small target detection accuracy in the prior art.

[0005] To achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0006] (1) Obtain training sample set and test sample set:

[0007] Obtain K two-dimensional images containing multiple small target categories, and mark the targets in each image. Then, the M images containing half of each small target category and their corresponding labels form the training sample set R1, and the remaining KM images and their corresponding labels form the test sample set E1, where K≥427.

[0008] (2) Construct a small target detection network model O based on the curvature attention mechanism:

[0009] Construct a small target detection network model O including a backbone network and a curvature network arranged in parallel, and an attention network connected to the output ends of the two networks; wherein the backbone network is used to obtain a segmentation map of each sample, and the backbone network includes N half-layer fusion blocks, each half-layer fusion block includes three half-layer feature extraction blocks arranged in parallel, each including a convolution layer, wherein the convolution layer in the second half-layer feature extraction block is further connected to two branches arranged in parallel, the first branch including a convolution layer, and the second branch including two convolution layer-nonlinear activation layers stacked in sequence; the curvature network is used to obtain an attention map of each sample; and the attention network is used to perform feature fusion on each segmentation map and its corresponding attention map, N ≥ 3;

[0010] (3) Initialization parameters:

[0011] The number of initialization iterations is s, the maximum number of iterations is S, S ≥ 10000, and the target detection network model O of the sth iteration s The weight parameter is w s , and let s = 0;

[0012] (4) Train the target detection network model O:

[0013] Randomly select L training samples from the training sample set R1 as the input of the target detection network model O for forward propagation to obtain the small target detection results of L training samples;

[0014] (5) Parameter w of the target detection network model s To update:

[0015] Through the detection results of L training samples, the w of the target detection network model is s Update and get the network model O of this iteration s ;

[0016] (6) Determine whether s≥S is true. If so, obtain the trained target detection network model O * Otherwise, let s=s+1, O=O s , and execute step (4);

[0017] (7) Obtaining the detection results of small targets:

[0018] The test sample set E is used as the trained target detection network model O * The input is forward propagated to obtain the small target detection results corresponding to all test samples.

[0019] Compared with the prior art, the present invention has the following advantages:

[0020] In the process of training the small target detection network model and obtaining the detection results of small targets, the backbone network of the present invention can extract small target features of various scales, wherein the half-layer fusion block reduces the loss of features in the feature fusion process of different depths, and enhances the integrity of the prediction results. The attention map of each sample obtained by the curvature network contains the accurate edge contour attention of the small target, which can supplement the fuzzy edge prediction results in the segmentation map. At the same time, the curvature combines the second-order gradient and the first-order gradient mathematically and reliably, so that the edge information contained in the attention map has strong reliability. Then, each segmentation map obtained by the backbone network is combined with each attention map obtained by the corresponding curvature network through the attention network for feature fusion, avoiding the defects of the existing technology network in paying insufficient attention to key edge features and incomplete extraction of target information. The test results show that the present invention can effectively improve the accuracy of small target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Flowchart for the implementation of the present invention;

[0022] Figure 2 A schematic diagram of the structure of a small target detection network model constructed in an embodiment of the present invention;

[0023] Figure 3 A schematic diagram of the structure of a backbone network used in an embodiment of the present invention;

[0024] Figure 4 This is a schematic diagram of the structure of a half-layer fusion block used in an embodiment of the present invention;

[0025] Figure 5 This is a schematic diagram of the structure of the curvature network and attention network used in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] Reference Figure 1 , the present invention comprises the following steps:

[0028] Step 1) Obtain training sample set and test sample set:

[0029] Obtain K two-dimensional images containing multiple small target categories, and mark the targets in each image. Then, the M images containing half of each small target category and their corresponding labels form the training sample set R1, and the remaining KM images and their corresponding labels form the test sample set E1, where K≥427.

[0030] In this embodiment, the training sample set and the test sample set are obtained through the NUAA-SIRST dataset, and K=427, M=325;

[0031] (2) Construct a small target detection network model O based on the curvature attention mechanism, whose structure is as follows: Figure 2 As shown:

[0032] Construct a small target detection network model O including a backbone network and a curvature network arranged in parallel, and an attention network connected to the output ends of the two networks;

[0033] The backbone network is used to obtain the segmentation map of each sample, and its structure is as follows Figure 3 As shown, it includes a Stem block, three residual blocks, three half-layer fusion blocks and a nonlinear activation layer; the input end of the first residual block is also connected to the input end of the third half-layer fusion block, the input end of the second residual block is also connected to the input end of the second half-layer fusion block, and the input end of the third residual block is also connected to the input end of the first half-layer fusion block. The output of the third half-layer fusion block is connected to the thirteenth nonlinear activation layer after upsampling operation; wherein, the Stem block includes a first convolutional layer, a first nonlinear activation layer, a second convolutional layer, a second nonlinear activation layer, a third convolutional layer, a third nonlinear activation layer, and a pooling layer connected in sequence; each residual block adopts the ResNet-20 structure, and each residual block includes two parallel residual branches, the first residual branch includes a fourth convolutional layer, a first normalization layer, a fourth nonlinear activation layer, a fifth convolutional layer, and a fifth nonlinear activation layer stacked in sequence, and the second residual branch includes a sixth convolutional layer and a second normalization layer stacked in sequence;

[0034] Structural reference of half-layer fusion block Figure 4, each half-layer fusion block includes three half-layer feature extraction blocks arranged in parallel, each including a convolution layer, wherein the convolution layer in the second half-layer feature extraction block is further connected to two branches arranged in parallel, the first branch includes a convolution layer, and the second branch includes two convolution layers-nonlinear activation layers stacked in sequence, and the outputs of the two branches are combined by cascading; when a matrix with a larger depth is input to the half-layer fusion block, an upsampling operation is also required; in the first half-layer fusion block, the first half-layer feature extraction block includes the thirteenth convolution layer, the second half-layer feature extraction block includes the fourteenth convolution layer and two branches, the first branch includes the fifteenth convolution layer, the second branch includes the sixteenth convolution layer, the eighth nonlinear activation layer, the seventeenth convolution layer and the ninth nonlinear activation layer stacked in sequence, and the third half-layer feature extraction block includes the nineteenth convolution layer;

[0035] The specific parameters are set as follows: the convolution kernel size of the first convolution layer is 3*3, and the step size is 2; the convolution kernel size of the second, third, fourth, fifth, fifteenth, sixteenth, and seventeenth convolution layers is 3*3, and the step size is 1; the convolution kernel size of the sixth, thirteenth, fourteenth, eighteenth, and nineteenth convolution layers is 1*1, and the step size is 1; all normalization layers use layer normalization; the interpolate method is used in the upsampling operation, and the output space size is the input image size; the first to twelfth nonlinear activation layers are all implemented using the ReLU function, and the thirteenth nonlinear activation layer is implemented using the Sigmoid function;

[0036] The structures of the curvature network and attention network are as follows Figure 5 As shown, where:

[0037] The curvature network is used to obtain an attention map containing the edge attention of small objects. It consists of a differential curvature extraction layer, three gated convolution blocks, and a nonlinear activation layer connected in sequence. The input of the nth half-layer fusion block in the backbone network is also connected to the input of its corresponding nth gated convolution block.

[0038] Each gated convolution block includes two residual blocks, a normalization layer and two convolution layers-nonlinear activation layers and an attention layer stacked therewith; the first gated convolution block includes the fourth residual block, the seventh normalization layer, the fifth residual block, the thirty-second convolution layer, the fourteenth nonlinear activation layer, the thirty-third convolution layer, the eighth normalization layer, the fifteenth nonlinear activation layer and the first attention layer stacked in sequence; the output of the third gated convolution block is connected to the twentieth nonlinear activation layer, and the specific parameters are set as follows: the settings of the fourth and fifth residual blocks are the same as the first, second and third residual blocks in the backbone network; the convolution kernel size of the thirty-second and thirty-third convolution layers is 1*1, and the step size is 1; the fourteenth nonlinear activation layer is implemented using the ReLU function, and the fifteenth and twentieth nonlinear activation layers are implemented using the Sigmoid function;

[0039] The attention network includes the fourth attention layer and the head block, where the head block includes the 38th convolutional layer, the 13th normalization layer, and the 21st nonlinear activation layer stacked in sequence. The specific parameters are set as follows: the convolution kernel size of the 38th convolutional layer is 3*3, and the stride is 1; the 13th normalization layer is layer normalization; and the 21st nonlinear activation layer is implemented using the ReLU function.

[0040] (3) Initialization parameters:

[0041] The number of initialization iterations is s, the maximum number of iterations is S, S ≥ 10000, and the target detection network model O of the sth iteration s The weight parameter is w s , and let s = 0;

[0042] (4) Train the target detection network model O:

[0043] Randomly select L training samples with replacement from the training sample set R1 as the input of the target detection network model O for forward propagation, and obtain the small target detection results of L training samples:

[0044] (4a) In this embodiment, L=32, the size of each input image is 512*512, and the number of channels is 3. The Stem block in the backbone network downsamples each image to obtain 32 high-level downsampled feature maps, whose size is 128*128 and the number of channels is 16. The first residual block extracts features from the high-level downsampled feature maps to obtain 32 downsampled feature maps with a depth of 2, whose size is 64*64 and the number of channels is 32. The second residual block extracts features from the downsampled feature maps with a depth of 2 to obtain 32 downsampled feature maps with a depth of 3, whose size is 32*32 and the number of channels is 64. The third residual block extracts features from the downsampled feature maps with a depth of 3 to obtain 32 downsampled feature maps with a depth of 4, whose size is 16*16 and the number of channels is 128.

[0045] (4b) The downsampled feature map with a depth of 4 is used as the upsampled feature map with a depth of 4, and is fused with the downsampled feature map with a depth of 3 through the feature fusion of the first half-layer fusion block to obtain 32 upsampled feature maps with a depth of 3, whose size is 32*32 and the number of channels is 64. The second half-layer fusion block fuses the upsampled feature map with a depth of 3 and the downsampled feature map with a depth of 2 to obtain 32 upsampled feature maps with a depth of 2, whose size is 64*64 and the number of channels is 32. The third half-layer fusion block fuses the high-level downsampled feature map with the upsampled feature map with a depth of 2 to obtain 32 high-level upsampled feature maps with a size of 128*128 and the number of channels is 16. After upsampling, the high-level upsampled feature map is activated by the thirteenth nonlinear activation layer to obtain 32 segmentation maps with a size of 512*512 and the number of channels is 16.

[0046] Among them, the feature fusion steps in the half-layer fusion block are:

[0047] (4b1) The nth half-layer fusion block upsamples the upsampled feature map with a depth of 3-n and concatenates it with the downsampled feature map with a depth of 2-n to obtain a joint feature matrix. The first half-layer feature extraction block extracts features from the joint feature matrix to obtain the first half-layer feature map.

[0048] (4b2) The second half-layer feature extraction block in the nth half-layer fusion block extracts features from the joint feature matrix to obtain a second half-layer feature map;

[0049] (4b3) The third half-layer feature extraction block in the nth half-layer fusion block extracts features from the feature maps of the 2-n layers to obtain a third half-layer feature map;

[0050] (4b4) The nth half-layer fusion block performs matrix summation on the feature maps of the 3-n layer, the first half-layer feature map, the second half-layer feature map, and the third half-layer feature map to obtain 32 upsampled feature maps with a depth of 2-n;

[0051] (4c) The 32 original input images are input into the curvature network. The differential curvature extraction layer performs differential curvature calculations on them to obtain 32 curvature information maps with a size of 512*512 and a channel number of 1. The specific method is:

[0052] According to differential geometry, the differential curvature extraction layer uses the differential curvature operator D to calculate the differential curvature of the feature map u to obtain the curvature feature map. The calculation is defined as:

[0053]

[0054] Among them, g x 、g y Represents the first-order gradient of u in the x and y directions, gxx 、g yy Represents the second-order gradient of u in the x and y directions, g xy It represents the second-order gradient of u in the oblique direction, and its calculation formula is:

[0055] g x =i*G x

[0056] g y =i*G y

[0057] g xx =i*G xx

[0058] g yy =i*G yy

[0059] g xy =i*G xy

[0060] Among them, G x , G y Represents the first-order gradient operator in the x and y directions, G xx , G yy Represents the second-order gradient operator in the x and y directions, G xy Represents the second-order gradient operator in the oblique direction, and its matrix value is:

[0061]

[0062]

[0063]

[0064]

[0065]

[0066] The first gated convolution block fuses the curvature information map with the upsampled feature map of depth 4 to obtain the first guided attention map. The second gated convolution block fuses the curvature information map with the upsampled feature map of depth 3 to obtain the second guided attention map. The third gated convolution block fuses the curvature information map with the upsampled feature map of depth 2 to obtain the third guided attention map. After the third curvature feature guided map is activated by the twentieth nonlinear activation layer, 32 attention maps are obtained. The size of all guided attention maps and attention maps is 512*512, and the number of channels is 1.

[0067] (4d) 32 segmentation maps and 32 attention maps are input into the attention network. The fourth attention layer fuses each segmentation map and its corresponding attention map through the attention mechanism to obtain a curvature-guided feature map. The attention layer fuses the edge contour features contained in the attention map into the segmentation map, thereby significantly improving the integrity of small target detection. The head block performs target detection on the curvature-guided feature map and obtains the target detection results of 32 training samples.

[0068] (5) Parameter w of the target detection network model s To update:

[0069] Through the detection results of L training samples, the w of the target detection network model is s Update and get the network model O of this iteration s :

[0070] (5a) Using Dice loss function L Dice , and pass the target detection result of the lth sample The label y corresponding to the training sample set l , calculate the loss value of the small target detection network model

[0071]

[0072] Among them, ∑ represents the sum operation, ∩ represents the intersection operation, and |·| represents the modulo operation;

[0073] (5b) Using Adagrad optimizer, and passing to w s The partial derivative of The parameters w of the network model s To update:

[0074]

[0075]

[0076] Among them, h s Indicates the cumulative effect of w s The square of the partial derivative, η represents the learning rate, and ⊙ represents the matrix dot product.

[0077] (6) Determine whether s≥S is true. If so, obtain the trained target detection network model O * Otherwise, let s=s+1, O=O s , and execute step (4);

[0078] (7) Obtaining the detection results of small targets:

[0079] The test sample set E is used as the trained target detection network model O * The input is forward propagated to obtain the small target detection results corresponding to all test samples.

[0080] The following simulation experiments further illustrate the technical effects of the present invention.

[0081] 1. Simulation conditions and content:

[0082] The hardware platform for the simulation experiment is: the processor is Intel(R) The simulation experiment was conducted on a Platinum 8255C CPU with a 2.5GHz clock speed, 86GB of memory, and an NVIDIA GeForce RTX 3090 graphics card. The software platform used for the simulation experiments was the Ubuntu 18.04 operating system, Python version 3.7, and PyTorch version 1.7.1. Experiments were conducted on the NUAA-SIRST dataset, which contains 427 images. For each dataset, the dataset was split into training and test sets in a ratio of 325:102.

[0083] The detection accuracy of the present invention and the existing "an infrared small target detection method based on context information" are compared and simulated, and the results are shown in Table 1.

[0084] 2. Analysis of simulation results:

[0085] In this experiment, the intersection over union (IoU), which is commonly used in target detection, is used as an indicator to measure the accuracy of small target detection. Its calculation method is:

[0086]

[0087] Referring to Table 1, the intersection-over-union ratio of the prior art on the NUAA-SIRST dataset is 74.37%, and the intersection-over-union ratio of the present invention on the NUAA-SIRST dataset is 78.76%. Compared with the prior art, the detection accuracy is significantly improved.

[0088] Table 1

[0089] method Existing technology The present invention Intersection and Union 74.37％ 78.76％

Claims

1. A small target detection method based on curvature attention mechanism, characterized in that: The following steps are involved: (1) Obtain training sample set and test sample set: Obtain K two-dimensional images containing multiple small target categories, and mark the targets in each image. Then, a total of M images containing more than half of each small target category and their corresponding labels form the training sample set R1, and the remaining KM images and their corresponding labels form the test sample set E1, where K ≥ 427. (2) Construct a small target detection network model O based on the curvature attention mechanism: Construct a small target detection network model O including a backbone network and a curvature network arranged in parallel, and an attention network connected to the output ends of the two networks; wherein the backbone network is used to obtain a segmentation map of each sample, and the backbone network includes N half-layer fusion blocks, each half-layer fusion block includes three half-layer feature extraction blocks arranged in parallel, each including a convolution layer, wherein the convolution layer in the second half-layer feature extraction block is further connected to two branches arranged in parallel, the first branch including a convolution layer, and the second branch including two convolution layer-nonlinear activation layers stacked in sequence; the curvature network is used to obtain an attention map of each sample; and the attention network is used to perform feature fusion on each segmentation map and its corresponding attention map, N ≥ 3; (3) Initialization parameters: The number of initialization iterations is s, the maximum number of iterations is S, S ≥ 10000, and the target detection network model O of the sth iteration s The weight parameter is w s , and let s = 0; (4) Train the target detection network model O: Randomly select L training samples from the training sample set R1 as the input of the target detection network model O for forward propagation to obtain the small target detection results of L training samples; (5) Parameter w of the target detection network model s To update: Through the detection results of L training samples, the w of the target detection network model is s Update and get the network model O of this iteration s ; (6) Determine whether s≥S is true. If so, obtain the trained target detection network model O * Otherwise, let s=s+1, O=O s , and execute step (4); (7) Obtaining the detection results of small targets: The test sample set E1 is used as the trained target detection network model O * The input is forward propagated to obtain the small target detection results corresponding to all test samples.

2. The method according to claim 1, characterized in that The small target detection network model O described in step (2), wherein: The backbone network further includes a Stem block, N residual blocks and a nonlinear activation layer connected in sequence, N half-layer fusion blocks are loaded between the N residual blocks and the nonlinear activation layer, and the input end of the nth half-layer fusion block is also connected to the input end of the N+1-nth residual block; wherein the Stem block includes three stacked composite layers consisting of sequentially stacked convolutional layers, normalization layers and nonlinear activation layers; the residual block includes two parallel residual branches, the first residual branch includes two sequentially stacked convolutional layers, a normalization layer and two nonlinear activation layers, and the second residual branch includes a convolutional layer and a normalization layer; The curvature network includes a differential curvature extraction layer, N gated convolution blocks, and a nonlinear activation layer connected in sequence, and the input of the nth half-layer fusion block is also connected to the input of its corresponding nth gated convolution block; the differential curvature extraction layer includes multiple gradient extraction operators; the gated convolution block includes two residual blocks, and a normalization layer, two convolution layers-nonlinear activation layers, and an attention layer stacked therewith; The attention network consists of an attention layer and a head block connected to it, where the head block includes a convolutional layer, a normalization layer, and a nonlinear activation layer stacked in sequence.

3. The method according to claim 1, characterized in that The target detection network model O is trained as described in step (4) and the implementation steps are as follows: (4a) The Stem block in the backbone network downsamples each sample, and N residual blocks extract features from each high-level downsampled feature map obtained by downsampling, obtaining L groups of downsampled feature maps of different depths; N half-layer fusion blocks fuse features from each group of downsampled feature maps, obtaining L groups of upsampled feature maps of different depths; the nonlinear activation layer activates the high-level upsampled feature maps, obtaining segmentation maps corresponding to the L training samples; (4b) The differential curvature extraction layer in the curvature network performs differential geometry calculations on each sample to obtain the curvature feature maps corresponding to L training samples; N gated convolution blocks perform feature fusion on each curvature feature map and its corresponding upsampled feature maps of different depths to obtain L curvature attention maps; The nonlinear activation layer activates each curvature attention map to obtain the attention maps corresponding to L training samples; (4c) The attention layer in the attention network performs attention fusion on each segmentation map and its corresponding attention map, and the head block performs target detection on the curvature-guided feature map obtained by attention fusion to obtain the target detection results of L training samples.

4. The method according to claim 3, characterized in that The downsampled feature maps of different depths described in step (4a) refer to each group including N downsampled feature maps, the depth of each downsampled feature map corresponds to the serial number of the extracted residual block, the downsampled feature map extracted by the first residual block is the highest layer downsampled feature map, and the downsampled feature map extracted by the nth residual block is a downsampled feature map with a depth of n.

5. The method according to claim 3, characterized in that The N half-layer fusion blocks described in step (4a) perform feature fusion on each set of downsampled feature maps respectively. The implementation steps are as follows: (4a1) The nth half-layer fusion block concatenates the upsampled feature map of depth 3-n with the downsampled feature map of depth 2-n, and extracts features from the joint feature matrix formed by the concatenation through the first half-layer feature extraction block in the nth half-layer fusion block to obtain the first half-layer feature map; (4a2) The second half-layer feature extraction block in the nth half-layer fusion block extracts features from the joint feature matrix to obtain a second half-layer feature map; (4a3) The third half-layer feature extraction block in the nth half-layer fusion block extracts features from the feature maps of the 2-n layers to obtain a third half-layer feature map; (4a3) The nth half-layer fusion block performs matrix summation on the feature maps of the 3-n layer, the first half-layer feature map, the second half-layer feature map, and the third half-layer feature map to obtain L upsampled feature maps with a depth of 2-n.

6. The method according to claim 3, characterized in that The curvature feature map corresponding to the L training samples described in step (4b), where the lth training sample u l The corresponding curvature feature map D(u l ) is calculated as: Among them, g l,x 、g l,y Represents u l The first-order gradient in the x and y directions, g l,xx 、g l,yy Indicates u l The second-order gradient in the x and y directions, g l,xy Indicates u l Second order gradient in oblique direction.

7. The method according to claim 5, characterized in that The parameter w of the target detection network model described in step (5) s To update, the steps are: (5a) Using Dice loss function L Dice , and pass the target detection result of the lth sample The label y corresponding to the training sample set l , calculate the loss value of the small target detection network model Among them, ∑ represents the sum operation, ∩ represents the intersection operation, and |·| represents the modulo operation; (5b) Using Adagrad optimizer, and passing to w s The partial derivative of For the model parameter w s To update: Among them, h s Indicates the cumulative effect of w s The square of the partial derivative, η represents the learning rate, and ⊙ represents the matrix dot product.

Images