Multi-scale forward feature gain infrared dim target detection method in complex environment

Abstract

The present invention discloses a multi-scale forward feature gain infrared small target detection method in complex environments. The method constructs and trains a multi-scale feature enhancement aggregation network, and utilizes the residual feature compensation module in the network to perform background suppression and target enhancement operations on infrared images. The method utilizes the selective feature fusion module in the multi-scale feature enhancement aggregation network to extract, learn, and fuse features of multiple scales in the infrared image. The present invention primarily addresses the problem of detecting small targets in infrared images under complex backgrounds, and has the advantages of effectively enhancing targets and suppressing backgrounds, fully extracting features of different scales to obtain information at all levels of infrared targets, and achieving high accuracy in detecting small infrared targets.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and further relates to a multi-scale forward feature gain infrared small target detection method in complex environments. The present invention uses a multi-scale forward feature gain method to detect infrared small targets, and can be used to detect infrared small targets in complex backgrounds. Background Art

[0002] Due to interference from complex backgrounds such as atmospheric clouds, infrared radiation, and target infrared interferometry, coupled with detection distance and the existence of certain anti-detection techniques, targets often appear as small-pixel objects lacking characteristic information such as shape, scale, and texture, and in extreme cases, may only occupy a single pixel. However, existing technologies cannot effectively utilize this contextual information, resulting in limited detection performance.

[0003] North University of China disclosed a method for detecting infrared weak and small targets based on multi-feature fusion in its patent application document "Infrared weak and small target detection method based on multi-feature fusion" (patent application number: 202111078520.0, application publication number CN113822352A). The specific steps for implementing this method are: in the first step, based on the property that the local grayscale value of the infrared weak and small target is large, the grayscale contrast between the target and the neighborhood background is used to enhance the real target and suppress part of the complex background. In the second step, the grayscale information of the infrared weak and small target itself conforms to the characteristic of the two-dimensional Gaussian distribution, and the target is detected by calculating the distance between the edge pixel point and the center pixel point of the infrared weak and small target and the covariance of the grayscale difference, and a significant difference is obtained. Figure 1 The third step is to use the low similarity between infrared weak targets and their neighbors to detect targets by calculating similarity factors, and obtain significant Figure 2 The fourth step will be to Figure 1 With significant Figure 2 A dot product is performed, and the final saliency map is obtained by integrating the various characteristics of the small infrared target. A simple threshold is calculated on this saliency map and segmented to obtain the final detection result. However, this method is only suitable for low-contrast images because it uses the grayscale contrast between the target and the surrounding background to enhance the true target and suppress parts of the complex background. High-contrast images will be over-enhanced and distorted using this method. Furthermore, the background suppression in the first step is partial, and the enhancement is limited to the local area of ​​the target and its surroundings, losing important information about the small infrared target.

[0004] The University of Electronic Science and Technology of China (USTC) has filed a patent application titled "A Method for Detecting Small Infrared Targets Based on Convolutional Neural Networks" (Patent Application No. 201710758679.4, Publication No. CN 107563433 A). The method involves the following steps: first, training the convolutional neural network with test data to determine network parameters; second, performing spectral residual (SR) on the infrared image to obtain candidate target regions; third, performing multi-scale sliding window detection on the candidate regions using the convolutional neural network to obtain preliminary classification results for the candidate regions; and fourth, performing non-maximum suppression on the preliminary classification results of the candidate regions, using the contours of all blocks with the largest overlap as the final target bounding box. However, this method has drawbacks: SR, which uses spectral residual (SR) to obtain candidate target regions, results in the loss of some low-level features in the spectral residual image. Subsequent multi-scale detection results in a loss of semantic detail in feature extraction, increasing the false alarm rate. The non-maximum suppression operation is likely to eliminate two closely related targets, leading to detection failure, and the lack of effective background suppression may result in interference from complex backgrounds. Summary of the Invention

[0005] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and propose a multi-scale forward feature gain infrared dim small target detection method in complex environments, which is used to solve the problem of distortion of images with high contrast due to excessive enhancement; the problem of losing important information containing small infrared targets; and the problem of increased false alarm rate due to loss of semantic details in feature extraction.

[0006] The technical approach to achieving the objectives of the present invention is to construct a multi-scale feature enhancement and aggregation network. This network comprises a residual feature compensation module, a selective feature fusion module, and a U-Net-based backbone network. The residual feature compensation module uses convolution kernels of varying sizes to collect features at different spatial scales. It directly subtracts the image under test from the feature-extracted image, and then enhances the required multi-scale features using linear scaling. This amplifies target features while suppressing background noise, effectively addressing the issues of prior art ineffective background suppression and distortion caused by over-enhancement in high-contrast images. The present invention constructs a selective feature fusion network that uses a low-medium refinement method to enable features at different levels to learn from each other. Specifically, it selectively utilizes features from the low, medium, and high spatial dimensions in different combinations for fusion. All three image features from the original image are extracted and mutually learned and fused, addressing the issue of missing important information containing small infrared targets. The present invention designs a U-Net-based backbone network to extract features at three different levels, low, medium, and high, from the target-enhanced and background-suppressed image, which has passed through the first sub-network. These features are then fed into the second sub-network. The backbone network connects all sub-modules in series, fully extracts features of different scales to obtain information of infrared targets at all levels while enhancing the target and suppressing the background, so as to solve the problem of losing semantic details in feature extraction and increasing the false alarm rate.

[0007] To achieve the above object, the specific implementation steps of the present invention include the following:

[0008] Step 1: Build a residual feature compensation module:

[0009] Step 1.1: Build a residual feature compensation module. The input layer, convolutional layer group, splicing unit, convolutional layer, subtraction unit, activation layer, multiplication unit, addition unit, and output layer of the module are connected in series in sequence. The input layer is connected to the subtraction unit and addition unit respectively. The convolutional layer group consists of two convolutional layers.

[0010] In step 1.2, set the channel parameter of the input layer to 16; set the convolution kernel size of the two convolution layers in the convolution layer group to 5×5 and 3×3 respectively, set the sliding step size to 1, and set the output channel to 16; set the convolution kernel size of the convolution layer in the residual feature compensation module to 1×1, and set the sliding step size to 1; use ReLU as the activation function in the activation layer; set the multiplication unit to multiply the image by the linear magnification factor β, where β is any real number selected in the range of [1,4];

[0011] Step 2: Build a selective feature fusion module:

[0012] Build a selective feature fusion module including the first input layer, the second input layer, the third input layer, the low-refinement submodule, the medium-refinement submodule, the addition unit, and the output layer; where:

[0013] The low-refinement submodule, addition unit, and output layer are connected in series in sequence; the low-refinement submodule is composed of an input group, a deconvolution group, a convolution group, and an inner product unit in series in sequence, and the input group is connected to the inner product unit; the input group is composed of a first input layer, a second input layer, and a third input layer; the convolution group is composed of three convolution layers; the deconvolution group is composed of two deconvolution layers; the channel parameters of the first to third input layers are set to 64, 32, and 16 respectively; the convolution kernel sizes of the two deconvolution layers in the deconvolution group are both set to 3×3, and the sliding step sizes are both set to 1; the convolution kernel sizes of the three convolution layers in the convolution layer group are both set to 3×3, and the sliding step sizes are both set to 1; the output channel is set to 16;

[0014] The mid-refining submodule is connected to the addition unit, and the first input layer is connected to the multiplication unit; the mid-refining submodule is composed of the second input layer, the pooling layer, the activation layer, the multiplication unit, and the deconvolution layer connected in series in sequence; the pooling window size of the pooling layer is set to 2×2, the sliding step size is set to 2, and the activation function of the pooling layer adopts ReLU; the activation function in the activation layer adopts the Sigmoid function; the convolution kernel size in the deconvolution layer is set to 3×3, and the sliding step size is set to 1;

[0015] Step 3: Construct the residual convolution group and residual deconvolution group:

[0016] Step 3.1: Build the first to third residual convolution groups. Each residual convolution group consists of four residual convolution units with the same structure and parameters connected in series. The structure of the residual convolution unit is as follows: the input layer, the first convolution layer, the first pooling layer, the second convolution layer, the addition unit, the second pooling layer, and the output layer are connected in series, and the input layer is also connected to the addition unit. The activation function of the first and second pooling layers uses ReLU. The convolution kernel size of the first and second convolution layers in the three residual convolution groups is set to 3×3, the sliding step size is set to 1, and the output channels are set to 16, 32, and 64 respectively.

[0017] Step 3.2: Build the first and second residual deconvolution groups. Each residual deconvolution group consists of four residual deconvolution units with the same structure and parameters connected in series. The structure of the residual deconvolution unit is as follows: the input layer, the first deconvolution layer, the first pooling layer, the second deconvolution layer, the addition unit, the second pooling layer, and the output layer are connected in series, and the input layer is also connected to the addition unit. The activation function of the first and second pooling layers adopts ReLU. The convolution kernel size of the first and second deconvolution layers in the two residual deconvolution groups is set to 3×3, the sliding step size is set to 1, and the output channels are set to 32 and 16 respectively.

[0018] Step 4: Build a multi-scale feature enhancement aggregation network:

[0019] Step 4.1, build a U-Net submodule including the input layer, the first residual convolution group, the second residual convolution group, the third residual convolution group, the first merging unit, the first residual deconvolution group, the second merging unit, the second residual deconvolution group, the first output layer, the second output layer, and the third output layer. Its structure is: the input layer, the first residual convolution group, the second residual convolution group, the third residual convolution group, the first merging unit, the second merging unit, and the second residual deconvolution group are connected in series in sequence; the first residual convolution group is also connected to the second merging unit, the second residual convolution group is also connected to the first merging unit, the third residual convolution group is also connected to the first output layer, the first residual deconvolution group is also connected to the second output layer, and the second residual deconvolution group is also connected to the third output layer;

[0020] In step 4.2, the input layer, convolutional layer, residual feature compensation module, U-Net submodule, selective feature fusion module, and output layer are connected in series to form a multi-scale feature enhancement aggregation network; the convolutional layer kernel size is set to 3×3, the sliding step size is set to 1, and the output channels are set to 16;

[0021] Step 5: Generate training set:

[0022] Step 5.1, select at least 400 infrared images to form a sample set;

[0023] Step 5.2: The infrared images in the sample set are randomly mirrored, randomly scaled, expanded and cropped, and Gaussian blurred in sequence;

[0024] Step 5.3: normalize each processed image and form a training set with all normalized samples;

[0025] Step 6: Train the multi-scale feature enhancement aggregation network:

[0026] The training set is input into the multi-scale feature enhancement aggregation network. The loss value of the multi-scale feature enhancement aggregation network is calculated using the intersection-over-union loss function after the selected infrared image is input into the multi-scale feature enhancement aggregation network. The network parameters are optimized using the gradient optimization algorithm of the stochastic gradient descent method. The weight value of the multi-scale feature enhancement aggregation network is iteratively updated until the intersection-over-union loss function of the network converges, thereby obtaining a trained multi-scale feature enhancement aggregation network.

[0027] Step 7: Detect small infrared targets:

[0028] Step 7.1, cropping and normalizing the infrared image to be detected in sequence;

[0029] In step 7.2, the normalized infrared image is input into the trained multi-scale feature enhancement aggregation network to detect infrared small targets in the image; the area composed of multiple pixels with a pixel value of 1 in the network output image is regarded as the detected small target area, and the pixel value of 0 represents the area where the small target is not detected.

[0030] Compared with the prior art, the present invention has the following advantages:

[0031] First, the residual feature compensation module designed in the present invention is used for preliminary processing of the image to enhance the target and suppress the background. The residual feature compensation part directly subtracts the image to be tested from the image with extracted features. The subtraction operation on the entire image can focus on the overall image to suppress the complex background, and the linear amplification factor of the key parameter of the enhanced target can be flexibly adjusted to prevent the high-contrast image from being over-enhanced and distorted. The global focus solves the problem of losing important information containing small infrared targets. This makes the image types applicable to the detection of the present invention more extensive, the information of the detected small infrared targets more complete, and the accuracy of the detected small infrared targets is improved.

[0032] Second, in the process of fusing features at different levels, the present invention maps mid- and high-level features to low-level features through autocorrelation fusion to obtain low-refined features, so that the low-level features fully learn the high-level feature information, avoiding the problem that the image of the spectral residual method causes some low-level scale features to be lost; because the mid-refined method multiplies the refined mid-level features by the high-level features on the right, the high-level features are effectively learned. The addition and fusion of the two refined features enable the present invention to fully learn and fuse features at different levels, prevent the loss of extracted semantic details, and reduce the false alarm rate.

[0033] Third, the multi-scale feature enhancement aggregation network of the present invention utilizes the U-Net algorithm to extract image features at different levels. Since the U-Net algorithm itself has performed preliminary extraction and fusion operations on features at different levels of the original image, the low, medium and high features input into the selective feature fusion module are further learned and fused. After two rounds of feature learning and fusion, multi-scale detection greatly avoids the loss of semantic details in feature extraction. It avoids the problem that the non-maximum suppression operation is likely to eliminate two similar targets, resulting in detection failure. The multi-level and multi-stage feature extraction gives the present invention the advantages of focusing on the information of each different small target and fully fusing features at different levels. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Flowchart for the implementation of the present invention;

[0035] Figure 2 Schematic diagram of the structure of the residual feature compensation module of the present invention;

[0036] Figure 3 Schematic diagram of the structure of the selective feature fusion module of the present invention;

[0037] Figure 4 Schematic diagram of the structure of the residual convolution group and the residual deconvolution group of the present invention;

[0038] Figure 5 Schematic diagram of the structure of the U-Net submodule and multi-scale feature enhancement aggregation network of the present invention;

[0039] Figure 6 Schematic diagram of an infrared image to be measured in an embodiment of the present invention;

[0040] Figure 7 Schematic diagram of the actual position and shape of a small infrared target in an infrared image to be measured according to an embodiment of the present invention;

[0041] Figure 8 Schematic diagram of the position and shape of a small infrared target detected in an infrared image to be detected in an embodiment of the present invention; DETAILED DESCRIPTION

[0042] The present invention is further described below with reference to the accompanying drawings and embodiments.

[0043] Reference Figure 1 The implementation steps of the present invention are further described in detail with reference to the accompanying drawings and embodiments.

[0044] Step 1: Generate training set and test set.

[0045] In step 1.1, the embodiment of the present invention selects 427 infrared images from the public SIRST data set to form a sample set, 80% of the images in the sample set form a training sample set, and 20% of the images form a test sample set.

[0046] In step 1.2, to ensure the training effect of the network, the infrared images in the training sample set are subjected to data augmentation processing. That is, each infrared image in the training sample set is subjected to random mirroring, random scaling, expansion and cropping, and Gaussian blur operations in sequence, so that each image is processed to a size of 480×480.

[0047] In step 1.3, each processed image is input into the torchvision module for normalization, and all normalized training samples are combined into a training set.

[0048] Step 2: Build a residual feature compensation module.

[0049] Reference Figure 2 , the structure of the residual feature compensation module constructed by the present invention is further described.

[0050] Step 2.1: Build a residual feature compensation module. The input layer, convolutional layer group, splicing unit, convolutional layer, subtraction unit, activation layer, multiplication unit, addition unit, and output layer in this module are connected in series in sequence. The input layer is connected to the subtraction unit and addition unit respectively. Its structure is that the convolutional layer group consists of two convolutional layers.

[0051] In step 2.2, set the parameters of the residual feature compensation module as follows.

[0052] The channel parameter of the input layer is set to 16; the convolution kernel sizes of the two convolution layers in the convolution layer group are set to 5×5 and 3×3, respectively, with a sliding step size of 1, and the output channel is set to 16; the convolution kernel size of the convolution layer in the residual feature compensation module is set to 1×1, with a sliding step size of 1. The activation function in the activation layer uses ReLU. The multiplication unit is set to multiply the image by a linear magnification factor β, where β is any real number selected in the range [1, 4]. In this embodiment of the present invention, β = 2.

[0053] Step 3: Construct a selective feature fusion module.

[0054] Reference Figure 3 , the structure of the selective feature fusion module constructed by the present invention is further described.

[0055] Build a selective feature fusion module including the first input layer, the second input layer, the third input layer, the low-refinement submodule, the medium-refinement submodule, the addition unit, and the output layer. Its structure is:

[0056] The low-refinement submodule, addition unit, and output layer are connected in series in sequence. The low-refinement submodule is composed of an input group, a deconvolution group, a convolution group, and an inner product unit connected in series in sequence, and the input group is connected to the inner product unit. The input group is composed of the first input layer, the second input layer, and the third input layer. The convolution group is composed of three convolution layers. The deconvolution group is composed of two deconvolution layers. The channel parameters of the first to third input layers are set to 64, 32, and 16, respectively. The convolution kernel size of the two deconvolution layers in the deconvolution group is set to 3×3, and the sliding step size is set to 1. The convolution kernel size of the three convolution layers in the convolution layer group is set to 3×3, and the sliding step size is set to 1; the output channel is set to 16.

[0057] The mid-refining submodule is connected to the addition unit, and the first input layer is connected to the multiplication unit. The mid-refining submodule consists of the second input layer, pooling layer, activation layer, multiplication unit, and deconvolution layer, connected in series. The pooling window size of the pooling layer is set to 2×2, the sliding step size is set to 2, and the activation function of the pooling layer is ReLU; the activation function of the activation layer is the Sigmoid function. The convolution kernel size of the deconvolution layer is set to 3×3, and the sliding step size is set to 1.

[0058] Step 4: Construct the residual convolution group and the residual deconvolution group.

[0059] Reference Figure 4 , the structures of the residual convolution group and the residual deconvolution group constructed by the present invention are further described.

[0060] In step 4.1, construct the first residual convolution group. The first residual convolution group consists of four residual convolution units with the same structure and parameters, connected in series. The structure of the residual convolution unit is as follows: the input layer, the first convolution layer, the first pooling layer, the second convolution layer, the addition unit, the second pooling layer, and the output layer are connected in series, with the input layer also connected to the addition unit. The activation function of the first and second pooling layers is ReLU. The convolution kernel size of the first and second convolution layers is set to 3×3, the sliding step size is set to 1, and the output channels are set to 16.

[0061] In step 4.2, construct the second residual convolution group. The second residual convolution group consists of four residual convolution units with the same structure and parameters, connected in series. The structure of the residual convolution unit is as follows: input layer, first convolution layer, first pooling layer, second convolution layer, addition unit, second pooling layer, and output layer, connected in series, with the input layer also connected to the addition unit. The activation function of the first and second pooling layers is ReLU. The convolution kernel size of the first and second convolution layers is set to 3×3, the sliding stride is set to 1, and the output channels are set to 32.

[0062] In step 4.3, construct the third residual convolution group. The third residual convolution group consists of four residual convolution units with the same structure and parameters, connected in series. The structure of the residual convolution unit is as follows: the input layer, the first convolution layer, the first pooling layer, the second convolution layer, the addition unit, the second pooling layer, and the output layer are connected in series, with the input layer also connected to the addition unit. The activation function of the first and second pooling layers is ReLU. The convolution kernel size of the first and second convolution layers is set to 3×3, the sliding step size is set to 1, and the output channels are set to 64.

[0063] In step 4.4, construct the first residual deconvolution group. The first residual deconvolution group consists of four residual deconvolution units with the same structure and parameters, connected in series. The structure of the residual deconvolution unit is as follows: input layer, first deconvolution layer, first pooling layer, second deconvolution layer, addition unit, second pooling layer, and output layer, connected in series, with the input layer also connected to the addition unit. The activation function of the first and second pooling layers is ReLU. The convolution kernel size of the first and second deconvolution layers is set to 3×3, the sliding step size is set to 1, and the output channels are set to 32.

[0064] In step 4.5, construct a second residual deconvolution group. The second residual deconvolution group consists of four residual deconvolution units with the same structure and parameters, connected in series. The structure of the residual deconvolution unit is as follows: an input layer, a first deconvolution layer, a first pooling layer, a second deconvolution layer, an addition unit, a second pooling layer, and an output layer, connected in series, with the input layer also connected to the addition unit. The activation function of the first and second pooling layers is ReLU. The convolution kernel size of the first and second deconvolution layers is set to 3×3, the sliding step size is set to 1, and the output channels are set to 16.

[0065] Step 5: Construct a multi-scale feature enhancement aggregation network.

[0066] Reference Figure 5 , further describes the structure of the U-Net submodule and the multi-scale feature enhancement aggregation network constructed by the present invention.

[0067] Step 5.1, build a U-Net submodule including the input layer, the first residual convolution group, the second residual convolution group, the third residual convolution group, the first merging unit, the first residual deconvolution group, the second merging unit, the second residual deconvolution group, the first output layer, the second output layer, and the third output layer. Its structure is: the input layer, the first residual convolution group, the second residual convolution group, the third residual convolution group, the first merging unit, the second merging unit, and the second residual deconvolution group are connected in series in sequence; the first residual convolution group is also connected to the second merging unit, the second residual convolution group is also connected to the first merging unit, the third residual convolution group is also connected to the first output layer, the first residual deconvolution group is also connected to the second output layer, and the second residual deconvolution group is also connected to the third output layer;

[0068] In step 5.2, the input layer, convolutional layer, residual feature compensation module, U-Net submodule, selective feature fusion module, and output layer are sequentially connected in series to form a multi-scale feature enhancement aggregation network. The convolutional layer kernel size is set to 3×3, the sliding step size is set to 1, and the output channels are set to 16.

[0069] Step 6: Train the multi-scale feature enhancement aggregation network:

[0070] The training set is input into the multi-scale feature enhancement aggregation network. The intersection-over-union loss function is used to calculate the loss value of the multi-scale feature enhancement aggregation network after the selected infrared image is input into the multi-scale feature enhancement aggregation network. The gradient optimization algorithm of the stochastic gradient descent method is used to optimize the network parameters, and the weight value of the multi-scale feature enhancement aggregation network is iteratively updated until the intersection-over-union loss function of the network converges, thereby obtaining a trained multi-scale feature enhancement aggregation network.

[0071] The intersection-over-union loss function is as follows:

[0072]

[0073] Among them, loss represents the infrared image input to the multi-scale feature enhancement aggregation network. After one round of iteration, the network loss value output is, ∑ represents the summation operation, N represents the total number of image samples in the training set, n represents the index of the image sample in the training set, and p n Represents the probability of each pixel in the predicted input infrared image belonging to the real target, r n Represents the category of each pixel in the input image, ∈ represents the correction factor, and the value of ∈ 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.

[0074] Step 7: Detect small infrared targets:

[0075] In step 7.1, each infrared image in the test set is cropped to 480×480 and normalized using the torchvision module to obtain each preprocessed infrared image.

[0076] In step 7.2, each preprocessed infrared image is fed into a trained multi-scale feature enhancement aggregation network to detect infrared small targets. The network outputs regions of the infrared image consisting of multiple pixels with a value of 1 as detected small targets, while pixels with a value of 0 represent regions where no small targets were detected.

[0077] The following is a combination of the embodiments of the present invention and Figure 6 、 7 , 8, further describe the present invention.

[0078] Figure 6 In this embodiment of the present invention, an original infrared image without any preprocessing is selected from the test set. The small target to be detected in this image is hidden in a complex cloud background and occupies very few pixels, making it difficult for even the human eye to observe. After preprocessing, the image is input into the multi-scale feature enhancement aggregation network constructed and trained by the present invention. The infrared small target detection result output by the network is shown in the figure below. Figure 7 shown. Figure 7 The weak white pixel in the center shows the detected Figure 6 The position of the small target is shown in the figure, and the shape of the detected small target is shown in the white box in the lower right corner.

[0079] Figure 8 The present invention tests the concentration and Figure 6 The true position and true shape of the infrared weak target corresponding to the infrared image to be measured. Figure 7The white weak pixel in the center is the real position information of the weak target in the infrared image to be measured, and the white box in the lower right corner shows the real shape of the weak target.

[0080] Will Figure 7 and Figure 8 By comparison, it is obvious that the position and shape information of the infrared dim target detected by the method of the present invention are very similar to the real information, indicating that the multi-scale feature enhancement aggregation network constructed and trained by the present invention has obvious effects of enhancing the target and suppressing the background, and the multi-scale feature extraction is accurate and the fusion effect is good, which greatly reduces the false alarm rate in the detection of infrared dim target.

Claims

1. A multi-scale forward feature gain infrared dim small target detection method in complex environments, characterized by: A multi-scale feature enhancement aggregation network is constructed and trained, and the residual feature compensation module in the network is used to perform background suppression and target enhancement operations on infrared images. A selective feature fusion module is used to extract, learn, and fuse features of multiple scales in infrared images. The detection method includes the following steps: Step 1: Build a residual feature compensation module: Step 1.1: Build a residual feature compensation module. The input layer, convolutional layer group, splicing unit, convolutional layer, subtraction unit, activation layer, multiplication unit, addition unit, and output layer of the module are connected in series in sequence. The input layer is connected to the subtraction unit and addition unit respectively. The convolutional layer group consists of two convolutional layers. In step 1.2, set the channel parameter of the input layer to 16; set the convolution kernel size of the two convolution layers in the convolution layer group to 5×5 and 3×3 respectively, set the sliding step size to 1, and set the output channel to 16; set the convolution kernel size of the convolution layer in the residual feature compensation module to 1×1, and set the sliding step size to 1; use ReLU as the activation function in the activation layer; set the multiplication unit to multiply the image by the linear magnification factor β, where β is any real number selected in the range of [1,4]; Step 2: Build a selective feature fusion module: Build a selective feature fusion module including the first input layer, the second input layer, the third input layer, the low-refinement submodule, the medium-refinement submodule, the addition unit, and the output layer; where: The low-refinement submodule, addition unit, and output layer are connected in series in sequence; the low-refinement submodule is composed of an input group, a deconvolution group, a convolution group, and an inner product unit in series in sequence, and the input group is connected to the inner product unit; the input group is composed of a first input layer, a second input layer, and a third input layer; the convolution group is composed of three convolution layers; the deconvolution group is composed of two deconvolution layers; the channel parameters of the first to third input layers are set to 64, 32, and 16 respectively; the convolution kernel sizes of the two deconvolution layers in the deconvolution group are both set to 3×3, and the sliding step sizes are both set to 1; the convolution kernel sizes of the three convolution layers in the convolution layer group are both set to 3×3, and the sliding step sizes are both set to 1; the output channel is set to 16; The mid-refining submodule is connected to the addition unit, and the first input layer is connected to the multiplication unit; the mid-refining submodule is composed of the second input layer, the pooling layer, the activation layer, the multiplication unit, and the deconvolution layer connected in series in sequence; the pooling window size of the pooling layer is set to 2×2, the sliding step size is set to 2, and the activation function of the pooling layer adopts ReLU; the activation function in the activation layer adopts the Sigmoid function; the convolution kernel size in the deconvolution layer is set to 3×3, and the sliding step size is set to 1; Step 3: Construct the residual convolution group and residual deconvolution group: Step 3.1: Build the first to third residual convolution groups. Each residual convolution group consists of four residual convolution units with the same structure and parameters connected in series. The structure of the residual convolution unit is as follows: the input layer, the first convolution layer, the first pooling layer, the second convolution layer, the addition unit, the second pooling layer, and the output layer are connected in series, and the input layer is also connected to the addition unit. The activation function of the first and second pooling layers uses ReLU. The convolution kernel size of the first and second convolution layers in the three residual convolution groups is set to 3×3, the sliding step size is set to 1, and the output channels are set to 16, 32, and 64 respectively. Step 3.2: Build the first and second residual deconvolution groups. Each residual deconvolution group consists of four residual deconvolution units with the same structure and parameters connected in series. The structure of the residual deconvolution unit is as follows: the input layer, the first deconvolution layer, the first pooling layer, the second deconvolution layer, the addition unit, the second pooling layer, and the output layer are connected in series, and the input layer is also connected to the addition unit. The activation function of the first and second pooling layers adopts ReLU. The convolution kernel size of the first and second deconvolution layers in the two residual deconvolution groups is set to 3×3, the sliding step size is set to 1, and the output channels are set to 32 and 16 respectively. Step 4: Build a multi-scale feature enhancement aggregation network: Step 4.1, build a U-Net submodule including the input layer, the first residual convolution group, the second residual convolution group, the third residual convolution group, the first merging unit, the first residual deconvolution group, the second merging unit, the second residual deconvolution group, the first output layer, the second output layer, and the third output layer. Its structure is: the input layer, the first residual convolution group, the second residual convolution group, the third residual convolution group, the first merging unit, the second merging unit, and the second residual deconvolution group are connected in series in sequence; the first residual convolution group is also connected to the second merging unit, the second residual convolution group is also connected to the first merging unit, the third residual convolution group is also connected to the first output layer, the first residual deconvolution group is also connected to the second output layer, and the second residual deconvolution group is also connected to the third output layer; In step 4.2, the input layer, convolutional layer, residual feature compensation module, U-Net submodule, selective feature fusion module, and output layer are connected in series to form a multi-scale feature enhancement aggregation network; the convolutional layer kernel size is set to 3×3, the sliding step size is set to 1, and the output channels are set to 16; Step 5: Generate training set: Step 5.1, select at least 400 infrared images to form a sample set; Step 5.2: The infrared images in the sample set are randomly mirrored, randomly scaled, expanded and cropped, and Gaussian blurred in sequence; Step 5.3: normalize each processed image and form a training set with all normalized samples; Step 6: Train the multi-scale feature enhancement aggregation network: The training set is input into the multi-scale feature enhancement aggregation network. The cross-entropy dice mixing loss function is used to calculate the loss value of the multi-scale feature enhancement aggregation network after the selected infrared image is input into the multi-scale feature enhancement aggregation network. The gradient optimization algorithm of the stochastic gradient descent method is used to optimize the network parameters. The weight value of the multi-scale feature enhancement aggregation network is iteratively updated until the cross-entropy dice mixing loss function of the network converges, thereby obtaining a trained multi-scale feature enhancement aggregation network. Step 7: Detect small infrared targets: Step 7.1, cropping and normalizing the infrared image to be detected in sequence; In step 7.2, the normalized infrared image is input into the trained multi-scale feature enhancement aggregation network to detect infrared small targets in the image; the area composed of multiple pixels with a pixel value of 1 in the network output image is regarded as the detected small target area, and the pixel value of 0 represents the area where the small target is not detected.

2. The method for detecting small targets in infrared images based on a multi-scale feature enhancement aggregation network according to claim 1, characterized in that: The cross-entropy dice hybrid loss function described in step 5 is as follows: Among them, loss represents the infrared image input to the multi-scale feature enhancement aggregation network. After one round of iteration, the network loss value output is, ∑ represents the summation operation, N represents the total number of image samples in the training set, n represents the index of the image sample in the training set, and p n Represents the probability of each pixel in the predicted input infrared image belonging to the real target, r n Represents the category of each pixel in the input image, ∈ represents the correction factor, and the value of ∈ 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.

Images