Small target detection method based on multi-scale feature compensation and gated enhancement

Abstract

The present invention discloses a small target detection method based on multi-scale feature compensation and gated enhancement, which mainly solves the problem of low detection accuracy of existing methods. The implementation steps are as follows: obtaining a training sample set and a test sample set; constructing a small target detection model based on multi-scale feature compensation and gated enhancement: including a convolution module, a transposition module, a gated enhancement module, and a prediction module connected in sequence, and a multi-scale feature compensation module loaded between the convolution module and the transposition module; iteratively training the small target detection model; and obtaining small target detection results of the test samples. In the present invention, the multi-scale feature compensation module is used to extract shallow and deep features before and after the downsampling operation, and uses shallow features to compensate for detail features of deep features, thereby improving the representation capability of multi-scale context. The gated enhancement module is used to enhance target features and suppress background noise, thereby improving the signal-to-noise ratio of the target and improving the accuracy of small target detection.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and in particular relates to a small target detection method based on multi-scale feature compensation and gated enhancement, which can be used in forest fire detection systems and precision guidance systems. Background Art

[0002] Small objects are generally defined as objects whose area accounts for between 0.08% and 0.58% of the entire image. These objects have low resolution, small pixels, low contrast, and low signal-to-noise ratio. Object detection algorithms struggle to fully extract their features, posing a challenge. Deep learning-based small object detection algorithms are constantly evolving, with research broadly categorized into anchor box optimization, data augmentation, network structure optimization, multi-scale feature fusion, and the incorporation of contextual information.

[0003] For example, Harbin University of Science and Technology proposed a small target detection method based on CenterNet improved multi-scale feature fusion in its patent application document "A small target detection method based on CenterNet improved multi-scale feature fusion" (patent application number: CN202211459753.X, application publication number CN115631400A). The backbone network of the original CenterNet model is replaced with a backbone network that can obtain high resolution to obtain high-quality feature maps. The high-resolution backbone network introduces a fine-grained multi-scale feature fusion module, which divides the input feature map into four groups of the same specifications. Except for the first group of feature maps, each group of feature maps will be added to the output of the previous group of feature maps to obtain the output, and then a convolution operation will be performed. After the convolution is completed, the outputs of the four groups of feature maps will be spliced ​​in the channel dimension, and then the coordinate attention mechanism will be introduced and the Dropout operation will be performed. The backbone network generates four feature maps of different scales, which are then integrated into feature maps of the same number of channels and size. These maps are then concatenated as input to the hierarchical feature fusion module, which assigns different weights to feature maps of different resolutions, adaptively adjusting the ratio of high and low resolutions to enhance the feature information required for detecting targets of different sizes. However, the feature information of small targets in this invention is easily lost during downsampling, and the signal-to-noise ratio of these features is low, resulting in lower detection accuracy. 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 multi-scale feature compensation and gated enhancement, which is used to solve the problem of low detection accuracy of existing small target detection methods.

[0005] To achieve the above object, the specific implementation steps of the present invention include the following:

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

[0007] Obtain K small target images and label the small target objects in each small target image. Then, combine the M small target images and their corresponding labels into a training sample set R1, and combine the remaining KM small target images and their corresponding labels into a test sample set E1, where K≥500.

[0008] (2) Construct a small target detection model O based on multi-scale feature compensation and gated enhancement:

[0009] A small target detection model O is constructed, which includes a convolution module, a transposition module, a gated enhancement module and a prediction module connected in sequence. The convolution module includes N convolution layers stacked in sequence. The transposition module includes N transposed convolution layers stacked in sequence. A multi-scale feature compensation module is loaded between the input and output ends of the nth convolution layer and the N-(n-1)th transposed convolution layer. The input ends of the second and third transposed convolution layers are connected to the input end of the gated enhancement module. The multi-scale feature compensation module includes three branches arranged in parallel. The first and second branches are composed of six branches connected in sequence. The residual convolution module consists of a convolution layer and six residual convolution modules connected in sequence, and the input end of the second branch is also connected to the input end of the third branch. The output ends of the second, fourth, and sixth residual convolution modules in the three branches are added and connected; the residual convolution module includes six convolution layers connected in sequence; the gated enhancement module includes two pooling layers arranged in parallel and a convolution layer and a nonlinear activation layer cascaded therewith; the prediction module includes a convolution layer, a normalization layer, a nonlinear activation layer, a random dropout layer, and a convolution layer connected in sequence; wherein N ≥ 3;

[0010] (3) Iteratively train the small target detection model O:

[0011] (3a) Initialize the number of iterations to t, the maximum number of iterations to T, T ≥ 1000, and the small target detection model O of the tth iteration t The weight and bias parameters in are w t 、b t , and let t = 0, O t =O;

[0012] (3b) L training samples are randomly selected with replacement from the training sample set R1 as the input of the small target detection model O for forward propagation, where 1≤L≤M:

[0013] (3b1) The N convolutional layers in the convolution module downsample each training sample layer by layer, obtaining L downsampled feature maps output by each convolutional layer;

[0014] (3b2) Each multi-scale feature compensation module uses the input of the corresponding convolution layer in the convolution module to which it is connected to perform feature compensation on the output of the convolution layer, thereby obtaining L feature compensation feature maps output by each multi-scale feature compensation module;

[0015] (3b3) The nth transposed convolutional layer in the transposed module upsamples the sum of the output of the connected multi-scale feature compensation module and the output of the n-1th transposed convolutional layer to obtain L upsampled feature maps output by each transposed convolutional layer;

[0016] (3b4) The gated enhancement module uses the input of the second transposed convolutional layer and the output of the Nth transposed convolutional layer to remove background information and enhance target features from the input of the third transposed convolutional layer, and obtains L feature maps with high signal-to-noise ratio output by the gated enhancement module;

[0017] (3b5) The prediction module predicts each feature map with a high signal-to-noise ratio and obtains L small target detection results;

[0018] (3c) Using the Dice loss function L Dice And the cross entropy loss function L CE , and calculate O by generating L small target detection results and their corresponding L labels t The loss value L t , and then calculate L by the chain rule t For the weight parameter ω t and bias parameter b t Partial derivative of and Finally, according to Right t 、b t Update and get the network model O of this iteration t ;

[0019] (3d) Determine whether t ≥ T holds true. If so, obtain the trained small target detection model O*. Otherwise, set t = t + 1 and execute step (3b);

[0020] (4) Obtain the small target detection results of the test sample:

[0021] The test sample set E1 is used as the input of the trained small target detection model O* for forward propagation to obtain the small target detection results corresponding to KM test samples.

[0022] Compared with the prior art, the present invention has the following advantages:

[0023] The small target detection model based on multi-scale feature compensation and gated enhancement constructed by the present invention includes a convolution module, a transposition module, a gated enhancement module and a prediction module connected in sequence, and a multi-scale feature compensation module loaded between the convolution module and the transposition module. In the process of training the model and obtaining small target detection results, the multi-scale feature compensation module is used to extract shallow and deep features before and after the downsampling operation, and use the shallow features to compensate the detail features of the deep features, so that the detail features are propagated to the deepest part of the network to compensate for the target features lost in the downsampling operation and improve the representation ability of the multi-scale context. The gated enhancement module is used to enhance the target features and suppress background noise, thereby improving the signal-to-noise ratio of the target. Experimental results show that the present invention can effectively improve the accuracy of small target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Flowchart for the implementation of the present invention;

[0025] Figure 2 A small target detection model based on multi-scale feature compensation and gated enhancement adopted in an embodiment of the present invention;

[0026] Figure 3 A multi-scale feature compensation module used in an embodiment of the present invention;

[0027] Figure 4 The residual convolution module used in the embodiment of the present invention;

[0028] Figure 5 This is the gate enhancement module used in the embodiment of the present invention. DETAILED DESCRIPTION

[0029] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] Reference Figure 1 , the present invention comprises the following steps:

[0031] (1) Obtain training sample set and test sample set:

[0032] Obtain 1000 small target images contained in the IRSTD-1k dataset, and label the small targets in each small target image. Then, 600 small target images and their corresponding labels form the training sample set R1, and the remaining 400 small target images and their corresponding labels form the test sample set E1;

[0033] (2) Construct a small target detection model O based on multi-scale feature compensation and gated enhancement:

[0034] Reference Figure 2, further describes the small target detection model based on multi-scale feature compensation and gated enhancement adopted in the embodiment of the present invention.

[0035] A small target detection model O is constructed, which includes a convolution module, a transposition module, a gated enhancement module and a prediction module connected in sequence; the convolution module includes N=3 convolution layers stacked in sequence; the transposition module includes N=3 transposed convolution layers stacked in sequence; the input and output ends of the first convolution layer are connected to the input end of the first multi-scale feature compensation module, the output end of the first multi-scale feature compensation module is added (element-wise add) to the output end of the second transposed convolution layer and then connected to the input end of the third transposed convolution layer, the input and output ends of the second convolution layer are connected to the input end of the second multi-scale feature compensation module, and the output end of the second multi-scale feature compensation module is added (element-wise add) to the output end of the first transposed convolution layer. add) and then connected to the input end of the second transposed convolution layer, the input and output ends of the third convolution layer are connected to the input end of the third multi-scale feature compensation module, and the output end of the third multi-scale feature compensation module is connected to the input end of the first transposed convolution layer; the input ends of the second transposed convolution layer and the third transposed convolution layer are also connected to the input end of the gate enhancement module; wherein the multi-scale feature compensation module includes three branches arranged in parallel, the input end of the second branch is also connected to the input end of the third branch, and the first branch includes a first residual convolution module, a second residual convolution module, a third residual convolution module, a fourth residual convolution module, a fifth residual convolution module connected in sequence. The residual convolution module, the sixth residual convolution module, the second branch includes the seventh residual convolution module, the eighth residual convolution module, the ninth residual convolution module, the tenth residual convolution module, the eleventh residual convolution module, and the twelfth residual convolution module connected in sequence, the third branch includes the fourth convolution layer, the thirteenth residual convolution module, the fourteenth residual convolution module, the fifteenth residual convolution module, the sixteenth residual convolution module, the seventeenth residual convolution module, and the eighteenth residual convolution module connected in sequence, and the output ends of the second residual convolution module, the eighth residual convolution module, and the fourteenth residual convolution module are also additively connected (element-wise add), the output ends of the fourth residual convolution module, the tenth residual convolution module and the sixteenth residual convolution module are further added (element-wise add), and the output ends of the sixth residual convolution module, the twelfth residual convolution module and the eighteenth residual convolution module are further added (element-wise add); the residual convolution module includes a fifth convolution layer, a sixth convolution layer, a seventh convolution layer, an eighth convolution layer, a ninth convolution layer and a tenth convolution layer connected in sequence; the gated enhancement module includes a first pooling layer and a second pooling layer arranged in parallel, and the output ends of the two pooling layers are channel-concat-ed and then sequentially connected to the eleventh convolution layer and the first nonlinear activation layer, and the middle layer input end is dot-multiplied with the output end of the first nonlinear activation layer and then added to the low layer input end (element-wise add);The prediction module includes the twelfth convolutional layer, the normalization layer, the second nonlinear activation layer, the random dropout layer and the thirteenth convolutional layer connected in sequence;

[0036] The specific parameters are set as follows: the input channel is 16; the convolution kernel size of the first, second and third convolution layers is 3*3, the step size is 2, the padding is 1, and the output channels are 32, 64 and 128 respectively; the convolution kernel size of the first, second and third transposed convolution layers is 4*4, the step size is 2, the padding is 1, and the output channels are 64, 32 and 16 respectively; the convolution kernel size of the fourth convolution layer in the first, second and third multi-level feature compensation blocks is 3*3, the step size is 2, the padding is 1, and the output channels are 64, 128 and 256 respectively; the fifth, sixth, seventh, eighth, ninth and tenth convolution layers are all atrous convolutions, and the convolution kernel size is 4*4. 3*3, the expansion rate of the fifth and sixth convolutional layers is 1, the expansion rate of the seventh and eighth convolutional layers is 2, and the expansion rate of the ninth and tenth convolutional layers is 3; the convolution kernel size of the eleventh convolutional layer is 3*3, the stride is 1, the padding is 1, and the output channel is 1, the convolution kernel size of the twelfth convolutional layer is 3*3, the stride is 1, the padding is 1, and the output channel is 4, and the convolution kernel size of the thirteenth convolutional layer is 1*1, the stride is 1, and the output channel is 1; the first pooling layer adopts maximum pooling, and the second pooling layer adopts average pooling; the first nonlinear activation layer adopts Sigmoid function, and the second nonlinear activation layer adopts ReLU function; the normalization layer adopts batch normalization; the random dropout layer sets the node retention probability to 0.5.

[0037] Reference Figure 3 , further describes the multi-scale feature compensation module used in the embodiment of the present invention.

[0038] Specifically, the multi-scale feature compensation module uses high-resolution features with rich details to guide feature aggregation, where high-resolution features can compensate for the details lost by downsampling and pooling operations, and low-resolution features retain the semantics of detection. e n-1 represents the shallow features of the first branch input of the nth multi-scale feature compensation module, whose input comes from a shallower network, f e n Represents the deep features of the second and third branches of the n-th multi-scale feature compensation module, whose input comes from a deeper network. Indicates f in the third branch e n Down-sampled features. Since the low-level feature map f e n-1 Contains the fine details of the target, so f en-1 Edge and texture information that are very suitable for network learning objectives; high-level features It has a large receptive field and contains rich high-level semantic features, so Very suitable for the network to distinguish targets from background noise; intermediate features f e n It contains some detailed features and has less background noise.

[0039] use represents the jth residual convolution module of the first branch of the nth multi-scale feature compensation module, represents the jth residual convolution module of the second branch of the nth multi-scale feature compensation module, represents the jth residual convolution module of the third branch of the nth multi-scale feature compensation module, represents the feature extraction of the first branch of the n-th multi-scale feature compensation module, represents the feature extraction of the second branch of the n-th multi-scale feature compensation module, represents the feature extraction of the third branch of the n-th multi-scale feature compensation module, which are defined as:

[0040]

[0041]

[0042]

[0043] The multi-scale feature compensation module continuously compensates the middle-level features by extracting features from different layers, using M n,q It represents the feature after the qth addition of the three branches of the nth multi-scale feature compensation module (element-wise add). This process can be described by the following equation:

[0044]

[0045]

[0046]

[0047] Among them, M n,1 It represents the combined features of the first three branches of the nth multi-scale feature compensation module, as the intermediate features after the first feature compensation, M n,2 It represents the combined features of the three branches of the second time of the n-th multi-scale feature compensation module, as the intermediate features after the second feature compensation, M n,3 Indicates the combined features of the three branches of the nth multi-scale feature compensation module for the third time, which is the output of the multi-scale feature compensation module. In the present invention, the shallow feature fe n-1 is the input of the nth convolutional layer, the deep feature f e n is the output of the nth convolutional layer. Here, the convolutional layer performs a downsampling operation. The multi-scale feature compensation module is used to extract shallow and deep features before and after the downsampling operation, and use shallow features to compensate for the detail features of the deep features. Therefore, the multi-scale feature compensation module compensates for the target features lost in the downsampling operation and propagates the detail features to the deepest part of the network.

[0048] Reference Figure 4 , further describing the residual convolution module used in the embodiment of the present invention.

[0049] Two 3*3 atrous convolutional layers, each with a dilated size of 1, 2, and 3, are sequentially connected in pairs to form a dilated convolutional block. There are three dilated convolutional blocks in total, with dilation rates set to 1, 2, and 3, respectively. These three atrous convolutional blocks are nested and can be viewed as multiple residual connections. Specifically, the fifth and sixth convolutional layers are paired, the seventh and eighth convolutional layers are paired, and the ninth and tenth convolutional layers are paired. The input of the fifth convolutional layer is element-wise added to the output of the sixth convolutional layer, the input of the seventh convolutional layer is element-wise added to the output of the eighth convolutional layer, and the input of the ninth convolutional layer is element-wise added to the output of the tenth convolutional layer, and then element-wise added to the inputs of the fifth, seventh, and ninth convolutional layers. This residual connection method of the atrous convolutional block can expand the receptive field to capture more multi-scale contextual information while maintaining the size of the image output feature map and preserving the resolution.

[0050] Reference Figure 5 , further describing the gating enhancement module used in the embodiment of the present invention.

[0051] Specifically, the gated enhancement module first uses the deep features to remove redundant information through the average pooling layer and the maximum pooling layer, retains the target features, and then passes through the convolution layer and the Sigmoid activation layer to obtain a deep pure target feature map with a value range of (0, 1). The deep pure target feature is then used as a gating unit to suppress most of the noise in the middle layer features, that is, the deep pure target feature map is multiplied element by element with the middle layer feature map to obtain the middle layer pure target feature. Finally, the shallow features with fine detail features are used to enhance the target features in the middle layer pure target feature map, that is, the middle layer pure target feature map is added element by element with the shallow feature map to obtain features with a high signal-to-noise ratio. This process can be expressed as:

[0052] F kernel =σ(Conv(Max(F high ),Avg(F high ))) (7)

[0053] F purity =F kernel ×F mid (8)

[0054] F HSNR =F purity +F low (9)

[0055] Where Max(·), Avg(·), Conv(·) and σ(·) represent the maximum pooling layer, average pooling layer, convolution layer and Sigmoid activation layer respectively. kernel represents the deep pure target feature, F high represents the deep features, F mid represents the middle-level features, F purity represents the middle-level pure target feature, F low Represents shallow features, F HSNR Represents features with a high signal-to-noise ratio. In the present invention, the shallow features are the features input to the second transposed convolutional layer, the middle features are the features input to the third transposed convolutional layer, and the deep features are the features output by the Nth transposed convolutional layer. In this embodiment of the present invention, the shallow features are the features input to the second transposed convolutional layer, the middle features are the features input to the third transposed convolutional layer, and the deep features are the features output by the third transposed convolutional layer.

[0056] (3) Iteratively train the small target detection model O:

[0057] (3a) Initialize the number of iterations to t, the maximum number of iterations is T = 2000, and the small target detection model O of the tth iteration t The weight and bias parameters in are w t 、b t , and let t = 0, O t =O;

[0058] (3b) 32 training samples are randomly selected with replacement from the training sample set R1 as the input of the small target detection model O for forward propagation:

[0059] (3b1) The first convolutional layer downsamples the training sample to obtain the downsampled feature map x1, the second convolutional layer downsamples the downsampled feature map x1 to obtain the downsampled feature map x2, and the third convolutional layer downsamples the downsampled feature map x2 to obtain the downsampled feature map x3;

[0060] (3b2) the first multi-scale feature compensation module performs feature compensation on the downsampled feature map x1 using the training sample to obtain a feature-compensated feature map x4; the second multi-scale feature compensation module performs feature compensation on the downsampled feature map x2 using the downsampled feature map x1 to obtain a feature-compensated feature map x5; the third multi-scale feature compensation module performs feature compensation on the downsampled feature map x3 using the downsampled feature map x2 to obtain a feature-compensated feature map x6;

[0061] (3b3) The first transposed convolution layer upsamples the feature compensation feature map x6 to obtain the upsampled feature map x7, and adds the upsampled feature map x7 and the feature compensation feature map x5 row by row element to obtain the sum feature map x8. The second transposed convolution layer upsamples the sum feature map x8 to obtain the upsampled feature map x9, and adds the upsampled feature map x9 and the feature compensation feature map x4 element by element to obtain the sum feature map x 10 , the third transposed convolutional layer adds the feature map x 10 Upsample and get the upsampled feature map x 11 ;

[0062] (3b4) The gate enhancement module uses the summed feature map x8 and the upsampled feature map x 11 Add the feature map x 10 Remove background information and enhance target features to obtain feature map x 12 , specifically, upsample the feature map x 11 Remove redundant information to obtain a deep pure target feature map x' 11 , deep pure target feature map x' 11 Sum sum feature map x 10 Perform element-by-element multiplication to obtain the middle-level pure target feature map x' 10 , the middle layer pure target feature map x' 10 Add the sum feature map x8 element by element to obtain a feature map x with a high signal-to-noise ratio 12 ;

[0063] (3b5) The prediction module is used to predict the feature map x with high signal-to-noise ratio 12 Make predictions and get 32 ​​small target detection results.

[0064] (3c) Using the Dice loss function L Dice And the cross entropy loss function L CE , and calculate O by generating 32 small target detection results and their corresponding 32 labels t The loss value L t , and then calculate L by the chain rule t For the weight parameter ω t and bias parameter b t Partial derivative of and Finally, according to Right t 、b t Update and get the network model O of this iteration t ;

[0065] (3d) Determine whether t≥2000 holds. If so, obtain the trained small target detection model O*. Otherwise, set t=t+1 and execute step (3b).

[0066] (4) Obtain the small target detection results of the test sample:

[0067] The test sample set E1 is used as the input of the trained small target detection model O* for forward propagation to obtain the small target detection results corresponding to 400 test samples.

[0068] The following is a description of the technical effects of the present invention in conjunction with simulation experiments.

[0069] Simulation conditions, content and result analysis:

[0070] The hardware platform for the simulation experiment is an Intel(R) Core i9-9900K CPU with a main frequency of 3.5GHz, 32GB of memory, and an NVIDIA GeForce RTX 2080Ti graphics card. The software platform for the simulation experiment is an Ubuntu 16.04 operating system, Python version 3.7, and PyTorch version 1.7.1.

[0071] The intersection-over-union (IoU) evaluation indicator is used to compare the detection accuracy of the patent document "A small target detection method based on improved multi-scale feature fusion of CenterNet" (patent application number: CN202211459753.X, application publication number: CN115631400A) and the present invention on the test samples. The intersection-over-union (IoU) of the small target detection results of the existing method is 61.75%, and the intersection-over-union (IoU) of the small target detection results of the present invention is 64.87%. Compared with the existing technology, the detection accuracy of the present invention is significantly improved.

Claims

1. A small target detection method based on multi-scale feature compensation and gated enhancement, characterized in that: The following steps are involved: (1) Obtain training sample set and test sample set: Obtain K small target images and label the small target objects in each small target image. Then, combine the M small target images and their corresponding labels into a training sample set R1, and combine the remaining KM small target images and their corresponding labels into a test sample set E1, where K≥500. (2) Construct a small target detection model O based on multi-scale feature compensation and gated enhancement: A small target detection model O is constructed, which includes a convolution module, a transposition module, a gated enhancement module, and a prediction module connected in sequence. The convolution module includes N convolution layers stacked in sequence. The transposition module includes N transposed convolution layers stacked in sequence. A multi-scale feature compensation module is loaded between the input and output of the nth convolution layer and the N-(n-1)th transposed convolution layer. The inputs of the second and third transposed convolution layers are connected to the input of the gated enhancement module. The multi-scale feature compensation module includes three branches arranged in parallel. The first and second branches are composed of six residual convolution modules connected in sequence. The third branch is composed of a convolution layer and six residual convolution modules connected in sequence. The input end of the second branch is also connected to the input end of the third branch. The output ends of the second, fourth, and sixth residual convolution modules in the three branches are added and connected. The residual convolution module includes six convolution layers connected in sequence. The gated enhancement module includes two pooling layers arranged in parallel and a convolution layer and a nonlinear activation layer cascaded therewith. The prediction module includes a convolution layer, a normalization layer, a nonlinear activation layer, a random dropout layer, and a convolution layer connected in sequence. Wherein, N ≥ 3. (3) Iteratively train the small target detection model O: (3a) Initialize the number of iterations to t, the maximum number of iterations to T, T ≥ 1000, and the small target detection model O of the tth iteration t The weight and bias parameters in are w t 、b t , and let t = 0, O t =O; (3b) L training samples are randomly selected with replacement from the training sample set R1 as the input of the small target detection model O for forward propagation, where 1≤L≤M: (3b1) The N convolutional layers in the convolution module downsample each training sample layer by layer, obtaining L downsampled feature maps output by each convolutional layer; (3b2) Each multi-scale feature compensation module uses the input of the corresponding convolution layer in the convolution module to which it is connected to perform feature compensation on the output of the convolution layer, thereby obtaining L feature compensation feature maps output by each multi-scale feature compensation module; (3b3) The nth transposed convolutional layer in the transposed module upsamples the sum of the output of the connected multi-scale feature compensation module and the output of the n-1th transposed convolutional layer to obtain L upsampled feature maps output by each transposed convolutional layer; (3b4) The gated enhancement module uses the input of the second transposed convolutional layer and the output of the Nth transposed convolutional layer to remove background information and enhance target features from the input of the third transposed convolutional layer, and obtains L feature maps with high signal-to-noise ratio output by the gated enhancement module; (3b5) The prediction module predicts each feature map with a high signal-to-noise ratio and obtains L small target detection results; (3c) Using the Dice loss function L Dice And the cross entropy loss function L CE , and calculate O by generating L small target detection results and their corresponding L labels t The loss value L t , and then calculate L by the chain rule t For the weight parameter ω t and bias parameter b t Partial derivative of and Finally, according to Right t 、b t Update and get the network model O of this iteration t ; (3d) Determine whether t ≥ T holds true. If so, obtain the trained small target detection model O*. Otherwise, set t = t + 1 and execute step (3b); (4) Obtain the small target detection results of the test sample: The test sample set E1 is used as the input of the trained small target detection model O* for forward propagation to obtain the small target detection results corresponding to KM test samples.

2. The small target detection method based on multi-scale feature compensation and gated enhancement according to claim 1, characterized in that: Calculation O described in step (3c) t The loss value L t , and the t 、b t To update, the calculation and update formulas are: L t =L Dice +L CE Among them, p l Represents the probability of each pixel in the predicted input small target image belonging to the label category, r l Represents the category of each pixel in the input image, ε represents the correction factor, which is any real number selected in the range of (0,0.1). Its function is to prevent the denominator of the fraction from being zero, ω t 、b t Indicates O t The weights and bias parameters of all learnable parameters, w t '、b t ' indicates ω t 、b t The update result of , α represents the learning rate.

Images