Infrared small target detection method and system

Through multi-scale convolution and sparse sampling attention mechanisms, small targets are detected in infrared images, and the problems of poor detection effect and large calculation volume in the existing technology are solved, and small infrared target detection with high accuracy and low error detection rate are achieved , suitable for edge device deployment.

CN117011669BInactive Publication Date: 2024-11-01ANHUI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310814159.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-07-03
Publication Date
2024-11-01
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing small object detection algorithms are poor in infrared images and have large calculations, making it difficult to effectively detect small objects in complex backgrounds.

Method used

The feature-related attention mechanism of multi-scale convolution and sparse sampling is adopted to extract feature information of different scales through convolution, and the sparse sampling attention mechanism is used to adapt the correlation characteristics, reducing calculation parameters, and improving detection accuracy and recall.

Benefits of technology

It improves the accuracy and recall of small and medium-sized infrared images, reduces the false detection rate, simplifies the model inference speed, and is suitable for edge device deployment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117011669B_ABST
    Figure CN117011669B_ABST
Patent Text Reader

Abstract

The present application provides an infrared small target detection method and system, and the infrared small target detection method includes: using convolution to downsample a single-channel image into multiple multi-channel feature maps of different scales; extracting detail features of the multi-channel feature map based on convolution operation, and outputting a first feature map; based on a feature association attention mechanism of sparse sampling, converting the first feature map into a second feature map; upsampling the second feature map using a bilinear difference method, and obtaining a third feature map; adding the third feature map, the second feature map and the first feature map matrix, and outputting a fused feature map; integrating the multi-channel features of the fused feature map, and converting the convolved features into binary classification probabilities of pixel points; based on a saturated activation function, screening pixel points that meet the target to be detected, so as to solve the technical problems of poor effect and large computational complexity of existing small target detection algorithms.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of machine vision technology, and in particular to a method and system for detecting small infrared targets. Background Art

[0002] Infrared small target detection under complex background is an important research direction in the field of image analysis and processing, and is widely used in urban smart transportation, logistics management, agricultural and forestry development, public safety, disaster relief deployment, etc. Since infrared small target images have high background complexity, small target size and blurred appearance, the information carried by the image is very limited, especially the target to be detected lacks appearance information such as texture, shape, color, etc. Therefore, it is difficult for traditional target detection algorithms to achieve the expected results when applied to small target detection.

[0003] In recent years, with the continuous development of deep learning theory and the growing demand from the industry, the research on small target detection technology has attracted more and more attention, but there is currently little research on algorithms specifically for small target detection.

[0004] Existing small target detection algorithms generally propose some improvements or optimization strategies based on general target detection methods. They enhance the robustness of model detection effects by increasing the number of network layers, designing a backbone network that can extract richer features, and complicating the feature fusion process, thereby improving the detection performance of small targets. However, these methods all have the disadvantage of high computational complexity. Summary of the invention

[0005] The present application provides an infrared small target detection method and system to solve the technical problems of poor effect and large amount of calculation of existing small target detection algorithms.

[0006] The present application provides an infrared small target detection method, comprising:

[0007] Use convolution to downsample a single-channel image into multiple multi-channel feature maps of different scales;

[0008] Extracting detail features of the multi-channel feature map based on a convolution operation, and outputting a first feature map;

[0009] Transforming the first feature map into a second feature map based on a feature association attention mechanism of sparse sampling;

[0010] Upsampling the second feature map by using a bilinear interpolation method, and obtaining a third feature map;

[0011] Adding the third feature map, the second feature map and the first feature map matrix, and outputting a fused feature map;

[0012] Integrate the multi-channel features of the fused feature map, and convert the convolved features into binary classification probabilities of pixel points;

[0013] Based on the saturation activation function, the pixels that meet the target to be tested are screened.

[0014] Optionally, the step of downsampling a single-channel image into a plurality of multi-channel feature maps of different scales by using convolution includes:

[0015] Use a convolution kernel to downsample a single-channel image into multiple multi-channel feature maps of different scales;

[0016] Extracting low-level features of the multi-channel feature map, wherein the low-level features include at least one of edge, line, corner, etc. level and position information;

[0017] Based on a normalization algorithm, normalizing the multi-channel feature map;

[0018] Based on a linear rectification activation function, a nonlinear transformation is performed on the multi-channel feature map, and the transformed multi-channel feature map is output.

[0019] Optionally, the step of extracting detail features of the multi-channel feature map based on a convolution operation and outputting a first feature map includes:

[0020] Reducing the number of channels of the multi-channel feature map using convolution;

[0021] Reducing the scale of the multi-channel feature map after convolution through a fully connected layer and extracting detail features;

[0022] Normalizing the multi-channel feature map and restoring the scale of the multi-channel feature map;

[0023] Restoring the number of channels of the multi-channel feature map using convolution;

[0024] The multi-channel feature map is normalized, and the first feature map is output.

[0025] Optionally, the step of converting the first feature map into the second feature map based on the feature association attention mechanism of sparse sampling includes:

[0026] Divide the first feature map and obtain S×S non-overlapping regions, where the size of the first feature map is X∈R H×W×C , the size of each of the regions is The feature map matrix after segmentation can be expressed as

[0027] Based on the linear projection, the tensor of query Q, the tensor of key K, and the tensor of value V are obtained. The linear projection is:

[0028] Q=X r W q ,K=X r W k ,V=X r W v

[0029] Where W q ,W k ,W v ∈R C×C Represent the projection weights of query, key, and value respectively,

[0030] Multiply the query matrix Q by the key matrix K and obtain a matrix of size S 2 ×S 2 The adjacency matrix L of

[0031] Take out the first k maximum values ​​of each row in the adjacency matrix L to form the first matrix

[0032] Multiply the first matrix, the query matrix Q and the key matrix K to obtain the second matrix K g And the third matrix V g ,

[0033] The second matrix K g Perform batch matrix multiplication with the query matrix Q and merge the channels of each small block matrix;

[0034] Based on the normalized exponential function, the numerical values ​​are normalized into probability distribution and a matrix A is obtained;

[0035] The matrix A and the third matrix V g Batch matrix multiplication is performed, and each small block matrix is ​​re-merged to output the second feature map.

[0036] Optionally, the step of upsampling the second feature map by using a bilinear interpolation method and obtaining a third feature map includes:

[0037] Compressing the number of channels of the second feature map;

[0038] The second feature map after normalized convolution;

[0039] Based on a linear rectification function, performing a nonlinear transformation on the second characteristic map;

[0040] The second feature map output by the linear rectification function is upsampled by using a bilinear interpolation method, and a third feature map is obtained.

[0041] Optionally, the step of integrating the multi-channel features of the fused feature map and converting the convolved multi-channel features into binary classification probabilities of pixel points includes:

[0042] Integrate the multi-channel features of the fused feature map and reduce the dimension of the feature tensor;

[0043] The fused feature map after normalized convolution;

[0044] Based on a linear rectification function, performing a nonlinear transformation on the fused feature map;

[0045] The number of channels of the fused feature map is compressed by convolution to obtain a single-channel feature map;

[0046] The multi-channel features of the fused feature map are integrated, and the convolved features are converted into binary classification probabilities of pixel points.

[0047] Optionally, the step of screening pixel points that meet the target to be measured based on the saturation activation function includes:

[0048] Based on the saturation activation function, the single-channel feature map is mapped to a probability space. The value of the pixel represents the binary classification probability of the point. The loss function of the saturation activation function adopts the loss of mean square error. The loss function is:

[0049]

[0050] where K is the number of image frames in the training batch, is the probability likelihood map generated by the labels, and is the probability likelihood graph obtained by the network;

[0051] Based on the binary probability likelihood function, the binary classification probability of the pixel point is obtained. The binary probability likelihood function is:

[0052]

[0053] Among them, the pixels with a probability greater than 0.5 are considered as targets to be measured, and the pixels with a probability less than 0.5 are considered as non-targets to be measured;

[0054] Output the probability likelihood map of the target distribution to be tested and process it into a binary image.

[0055] Optionally, before the step of downsampling a single-channel image into a plurality of multi-channel feature maps of different scales by using convolution, the step further includes:

[0056] Unify the input images into single-channel images of a preset scale;

[0057] When the scale of the image is larger than the preset scale, the middle area of ​​the image is cropped. When the scale of the image is smaller than the preset scale, the missing area is filled with a constant value of 0.

[0058] Correspondingly, the present application also provides an electronic device, including a memory and a processor, the memory being used to store executable program code; the processor being connected to the memory, and running a computer program corresponding to the executable program code by reading the executable program code, so as to execute the steps in any of the above-mentioned infrared small target detection methods.

[0059] The present application also provides an infrared small target detection system, which includes an electronic device.

[0060] The present application provides an infrared small target detection method and system. The infrared small target detection method is based on a combination of multi-scale convolution and sparse attention mechanism, which can effectively improve the accuracy of small target detection in infrared images and reduce the false detection rate. It can play an important role in the field of infrared image target detection and has broad application prospects and economic benefits. Multi-scale convolution is used to extract feature information of different scales, and sparse sampling attention mechanism is combined to adaptively associate features of different scales and reduce calculation parameters, thereby achieving accurate positioning and detection of small targets.

[0061] Through multi-scale convolution downsampling, feature information of different scales is extracted, and the information of different scales is fused one by one after multi-step processing through the network, thus solving the defect that a single-scale convolutional network cannot efficiently extract the features of the target to be detected in small target detection, improving the detection accuracy and recall rate; at the same time, overcoming the problem of the small number of target features in the process of infrared small target detection.

[0062] By introducing a feature-association attention mechanism based on sparse sampling into the small-scale, multi-channel deep network in the multi-scale convolutional network, the network can remove irrelevant information during learning and focus on extracting small targets and surrounding environment and texture features, so that the final model reasoning speed is accelerated. This solves the problem of slow model reasoning speed caused by too many deep network parameters when using deep learning networks for small target detection.

[0063] The feature association attention mechanism based on sparse sampling can reduce the number of model parameters while retaining the features of the target to be measured, which reduces the model's computing power requirements and facilitates the deployment of edge devices. This solves the problem that deep learning networks are usually difficult to deploy on edge devices due to the large number of model parameters and high computing power requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0065] Figure 1a is a schematic diagram of cropping images in the infrared small target detection method provided in this application;

[0066] Figure 1b It is a schematic diagram of filling a picture in the infrared small target detection method provided by this application;

[0067] Figure 2 It is a flowchart of the infrared small target detection method provided by this application;

[0068] Figure 3 It is a flowchart of step S200 in the infrared small target detection method provided by the present application;

[0069] Figure 4 It is a flowchart of step S300 in the infrared small target detection method provided by the present application;

[0070] Figure 5 It is a flowchart of step S400 in the infrared small target detection method provided by the present application;

[0071] Figure 6 It is a flow chart of step S500 in the infrared small target detection method provided by the present application;

[0072] Figure 7 It is a flowchart of step S700 in the infrared small target detection method provided in this application. DETAILED DESCRIPTION

[0073] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application. In addition, it should be understood that the specific implementation methods described herein are only used to illustrate and explain the present application and are not used to limit the present application. In the present application, unless otherwise stated, the directional words used, such as "up", "down", "left", and "right", generally refer to the up, down, left, and right of the device in actual use or working state, specifically the drawing direction in the accompanying drawings.

[0074] The present application provides an infrared small target detection method and system, which are described in detail below. It should be noted that the description order of the following embodiments does not limit the preferred order of the embodiments of the present application. In the following embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0075] See also Figure 2 The present application provides an infrared small target detection method, which extracts the target's salient features through a multi-scale convolutional network and detects infrared small targets through a feature association attention mechanism based on sparse sampling. This method fully utilizes the underlying high-resolution features rich in detail information through multi-scale feature fusion, and performs sparse sampling on the feature map of the deep network for feature association, thereby reducing the calculation parameters and improving the inference speed of the model, and has good application prospects.

[0076] See also Figure 2 , the infrared small target detection method specifically includes the following steps:

[0077] S100, unifying the input images into single-channel images of a preset scale;

[0078] Input images of different sizes need to be unified into single-channel images of the same scale (512×512) to facilitate the design and unification of network parameters.

[0079] In order to keep the information contained in the picture from being lost as much as possible, when the scale of the picture is larger than the preset scale (512×512 in this embodiment), the middle area of ​​the picture can be cropped; when the scale of the picture is smaller than the preset scale, the missing area can be filled with a constant value of 0.

[0080] S200, downsampling a single-channel image into multiple multi-channel feature maps of different scales using convolution;

[0081] Use convolution to downsample a (512×512×1) single-channel image into multi-channel feature maps of different scales of (512×512×32), (256×256×64), (128×128×128), and (64×64×256) using a (3×3) convolution kernel.

[0082] The above step S200 specifically includes the following steps:

[0083] S210, down-sampling a single-channel image into a plurality of multi-channel feature maps of different scales using a convolution kernel;

[0084] Use convolution to downsample a (512×512×1) single-channel image into multi-channel feature maps of different scales of (512×512×32), (256×256×64), (128×128×128), and (64×64×256) using a (3×3) convolution kernel.

[0085] S220, extracting low-level features of the multi-channel feature map, wherein the low-level features include at least one of edge, line, corner, and other level and position information;

[0086] S230, performing normalization processing on the multi-channel feature map based on a normalization algorithm;

[0087] Normalizing the multi-channel feature maps makes the model easier to train.

[0088] S240, performing a nonlinear transformation on the multi-channel feature map based on a linear rectification activation function, and outputting the transformed multi-channel feature map;

[0089] The linear rectification function (ReLU) is used to perform nonlinear transformation on the multi-channel feature map to improve the expression ability and classification performance of the model.

[0090] S300, extracting detail features of the multi-channel feature map based on a convolution operation, and outputting a first feature map;

[0091] The above step S300 specifically includes the following steps:

[0092] S310, using convolution to reduce the number of channels of the multi-channel feature map;

[0093] A (1×1) convolution kernel is used to perform a convolution operation on the multi-channel feature map to reduce the number of channels of the multi-channel feature map and to reduce subsequent calculation parameters.

[0094] S320, reducing the scale of the multi-channel feature map after convolution through a fully connected layer, and extracting detail features;

[0095] The convolutional multi-channel feature map is reduced to a (512×512×C2) scale by a fully connected layer to extract the detailed features of the target.

[0096] S330, standardizing the multi-channel feature map, and restoring the scale of the multi-channel feature map;

[0097] The multi-channel feature map is standardized to make the model easier to train. The input (64×C2) scale multi-channel feature map is restored to the scale of (512×512×C2). The main function of this step is to restore the multi-channel feature map to its original size to amplify the detailed features extracted in the previous step.

[0098] S340, restoring the number of channels of the multi-channel feature map by using convolution;

[0099] The multi-channel feature map is convolved using a (1×1) convolution kernel. The main function of this step is to restore the number of channels of the multi-channel feature map to the original C1, so as to facilitate the subsequent fusion of feature maps of different scales.

[0100] S350, standardizing the multi-channel feature map, and outputting the first feature map;

[0101] Normalizing multi-channel feature maps makes the model easier to train.

[0102] This embodiment extracts feature information of different scales through multi-scale convolution downsampling, and fuses the information of different scales one by one after multi-step processing through the network, thereby solving the defect that a single-scale convolutional network cannot efficiently extract the features of the target to be detected in small target detection, and improving the detection accuracy and recall rate.

[0103] S400, transforming the first feature map into a second feature map based on a feature association attention mechanism of sparse sampling;

[0104] Since the detected object is a small target, in order to minimize the calculation parameters while keeping the features of the detected target intact, this embodiment adopts a feature association attention mechanism based on sparse sampling.

[0105] Step S400 specifically includes the following steps:

[0106] S410, dividing the first feature map and obtaining S×S non-overlapping regions, the size of the first feature map is X∈R H×W×C , the size of each of the regions is The feature map matrix after segmentation can be expressed as

[0107] S420. Based on linear projection, obtain the tensor of the query Q, the tensor of the key K, and the tensor of the value V. The linear projection is:

[0108] Q=X r W q ,K=X r W k ,V=X r W v

[0109] Where W q ,W k ,W v ∈R C×C Represent the projection weights of query, key, and value respectively,

[0110] S430, multiply the query matrix Q by the key matrix K, and obtain a matrix of size S 2 ×S 2 The adjacency matrix L of

[0111] S440, extract the first k maximum values ​​of each row in the adjacency matrix L to form a first matrix

[0112] S450, multiply the first matrix, the query matrix Q and the key matrix K to obtain a second matrix K g And the third matrix V g ,

[0113] S460: transform the second matrix K g Perform batch matrix multiplication with the query matrix Q and merge the channels of each small block matrix;

[0114] S470, based on a normalized exponential function, normalize the numerical value into a probability distribution, and obtain a matrix A;

[0115] S480, the matrix A and the third matrix V g Perform batch matrix multiplication, and re-merge each small block matrix to output the second feature map;

[0116] This embodiment introduces a feature association attention mechanism based on sparse sampling in the small-scale, multi-channel deep network in the multi-scale convolutional network, so that the network can remove irrelevant information and concentrate on extracting small targets and surrounding environment and texture features during learning, so that the final model reasoning speed is accelerated. This solves the problem in the prior art that when using deep learning networks for small target detection, the large number of deep network parameters leads to slow model reasoning speed.

[0117] S500, upsampling the second feature map by using a bilinear interpolation method, and obtaining a third feature map;

[0118] Step S500 specifically includes the following steps:

[0119] S510, compressing the number of channels of the second feature map;

[0120] The (1x1) convolution kernel is used to perform channel compression on the second feature map processed by the feature extraction method and the feature association attention method, and the number of channels is compressed from 2C to C. The main function of this step is to reduce the number of parameters and the amount of calculation in the model to increase the training speed and reduce the risk of overfitting.

[0121] S520, normalizing the second feature map after convolution;

[0122] Normalizing the second feature map after convolution makes the model easier to train.

[0123] S530: performing a nonlinear transformation on the second characteristic graph based on a linear rectification function;

[0124] The second feature map is nonlinearly transformed using the linear rectification function (ReLU) to improve the expressiveness and classification performance of the model.

[0125] S540, upsampling the second feature map output by the linear rectification function by using a bilinear difference method, and obtaining a third feature map;

[0126] The second feature map output by the linear rectification function is upsampled using bilinear interpolation to restore its size to twice its original size for subsequent feature fusion.

[0127] S600, adding the third feature map, the second feature map and the first feature map matrix, and outputting a fused feature map;

[0128] The upsampled third feature map and the shallow feature map are matrix-added to generate the final fused feature map. The main function of this step is to fuse the deep features with the shallow features to improve the classification performance of the model.

[0129] S700, integrating the multi-channel features of the fused feature map, and converting the convolved features into binary classification probabilities of pixel points;

[0130] The binary classification probability of each pixel is output through the saturated activation function (Sigmoid) to achieve the segmentation effect of small targets.

[0131] Step S700 specifically includes the following steps:

[0132] S710, integrating the multi-channel features of the fused feature map and reducing the dimension of the feature tensor;

[0133] Use a (3x3) convolution kernel to perform a convolution operation on the fused feature map, integrate the multi-channel features of the feature map, and perform the first dimensionality reduction on the feature tensor.

[0134] S720, normalizing the fused feature map after convolution;

[0135] The fused feature map after convolution is standardized to make the model easier to train. The batch normalization algorithm can effectively solve the problems of gradient disappearance and gradient explosion in deep neural networks and improve the training speed and accuracy of the model.

[0136] S730, performing a nonlinear transformation on the fused feature map based on a linear rectification function;

[0137] The linear rectification function (ReLU) is used to perform nonlinear transformation on the fused feature map to improve the expression ability and classification performance of the model.

[0138] S740, compressing the number of channels of the fused feature map by convolution to obtain a single-channel feature map;

[0139] The (1x1) convolution kernel is used again to perform a convolution operation on the fused feature map, compressing the number of channels from 16 to 1.

[0140] S750, integrating the multi-channel features of the fused feature map, and converting the convolved features into binary classification probabilities of pixel points.

[0141] S800, based on the saturation activation function, screening pixels that meet the target to be measured;

[0142] The single-channel feature map output in the previous step is mapped to the probability space of [0,1] using the saturation activation function (Sigmoid). The value of the pixel represents the binary classification probability of the point. The output of the model is the probability likelihood map of the target to be tested, where the closer the probability is to 1, the more likely the pixel is to belong to the target to be tested, and the closer the pixel value is to 0, the more likely the pixel is to belong to a non-target to be tested.

[0143] Step S800 specifically includes the following steps:

[0144] S810, based on the saturation activation function, the single channel feature map is mapped to a probability space, the value of the pixel point represents the binary classification probability of the point, and the loss function of the saturation activation function adopts the loss of mean square error, and the loss function is:

[0145]

[0146] where K is the number of image frames in the training batch, is the probability likelihood map generated by the labels, and is the probability likelihood graph obtained by the network;

[0147] S820, based on the binarization probability likelihood function, obtain the binary classification probability of the pixel point, the binarization probability likelihood function is:

[0148]

[0149] Among them, the pixels with a probability greater than 0.5 are considered as targets to be measured, and the pixels with a probability less than 0.5 are considered as non-targets to be measured;

[0150] S830, outputting a probability likelihood graph of the target distribution to be measured, and processing it into a binary image;

[0151] The final output of the model is the probability likelihood map of the target distribution. The probability of non-target areas (background, noise interference) is 0, so the use of binary probability likelihood function can further reduce the false detection rate.

[0152] The infrared small target detection method based on multi-scale convolution and sparse attention mechanism in this embodiment can effectively improve the accuracy and recall rate of small target detection in infrared images. This method uses multi-scale convolution to extract feature information of different scales, and combines sparse sampling attention mechanism to adaptively associate features of different scales and reduce calculation parameters, thereby achieving accurate positioning and detection of small targets.

[0153] Compared with traditional target detection methods, infrared small target detection methods have higher detection accuracy and lower false detection rate. They can play an important role in the field of infrared image target detection and have broad application prospects and economic benefits.

[0154] This embodiment extracts feature information of different scales through multi-scale convolution downsampling, and fuses the information of different scales one by one after multi-step processing through the network, thereby solving the defect that a single-scale convolutional network cannot efficiently extract the features of the target to be detected in small target detection, and improves the detection accuracy and recall rate; at the same time, it overcomes the problem of the small number of target features in the process of infrared small target detection.

[0155] By introducing a feature-association attention mechanism based on sparse sampling into the small-scale, multi-channel deep network in the multi-scale convolutional network, the network can remove irrelevant information during learning and focus on extracting small targets and surrounding environment and texture features, so that the final model reasoning speed is accelerated. This solves the problem of slow model reasoning speed caused by too many deep network parameters when using deep learning networks for small target detection.

[0156] The feature association attention mechanism based on sparse sampling can reduce the number of model parameters while retaining the features of the target to be measured, which reduces the model's computing power requirements and facilitates the deployment of edge devices. This solves the problem that deep learning networks are usually difficult to deploy on edge devices due to the large number of model parameters and high computing power requirements.

[0157] The present application also provides an infrared small target detection system, which includes an electronic device, which includes a memory and a processor; wherein the memory is used to store executable program code, and the processor is connected to the memory, and runs a computer program corresponding to the executable program code by reading the executable program code to execute the steps in the aforementioned infrared small target detection method.

[0158] The above is a detailed introduction to the infrared small target detection method and system provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A method for detecting small infrared targets, characterized in that: include: Use convolution to downsample a single-channel image into multiple multi-channel feature maps of different scales; Extracting detail features of the multi-channel feature map based on a convolution operation, and outputting a first feature map; Transforming the first feature map into a second feature map based on a feature association attention mechanism of sparse sampling; Upsampling the second feature map by using a bilinear interpolation method, and obtaining a third feature map; Adding the third feature map, the second feature map and the first feature map matrix, and outputting a fused feature map; Integrate the multi-channel features of the fused feature map, and convert the convolved features into binary classification probabilities of pixel points; Based on the saturation activation function, the pixels that meet the target to be tested are screened. The step of converting the first feature map into the second feature map based on the feature association attention mechanism of sparse sampling includes: Divide the first feature map and obtain S×S non-overlapping regions, where the size of the first feature map is X∈R H ×W×C , the size of each of the regions is The feature map matrix after segmentation can be expressed as Based on the linear projection, the tensor of query Q, the tensor of key K, and the tensor of value V are obtained. The linear projection is: Q=X r W q ,K=X r W k ,V=X r W v Where W q ,W k ,W v ∈R C×C Represent the projection weights of query, key, and value respectively, Multiply the query matrix Q by the key matrix K and obtain a matrix of size S 2 ×S 2 The adjacency matrix L of Take out the first k maximum values ​​of each row in the adjacency matrix L to form the first matrix Multiply the first matrix, the query matrix Q and the key matrix K to obtain the second matrix K g And the third matrix V g , The second matrix K g Perform batch matrix multiplication with the query matrix Q and merge the channels of each small block matrix; Based on the normalized exponential function, the numerical values ​​are normalized into probability distribution and a matrix A is obtained; The matrix A and the third matrix V g Perform batch matrix multiplication, and merge each small block matrix again to output the second feature map, The step of screening pixels that meet the target to be measured based on the saturation activation function includes: Based on the saturation activation function, the single-channel feature map is mapped to a probability space. The value of the pixel represents the binary classification probability of the point. The loss function of the saturation activation function adopts the loss of mean square error. The loss function is: where K is the number of image frames in the training batch, is the probability likelihood map generated by the labels, and is the probability likelihood graph obtained by the network; Based on the binary probability likelihood function, the binary classification probability of the pixel point is obtained. The binary probability likelihood function is: Among them, the pixels with a probability greater than 0.5 are considered as targets to be measured, and the pixels with a probability less than 0.5 are considered as non-targets to be measured; Output the probability likelihood map of the target distribution to be tested and process it into a binary image.

2. The infrared small target detection method according to claim 1, characterized in that: The step of downsampling a single-channel image into a plurality of multi-channel feature maps of different scales by using convolution comprises: Use a convolution kernel to downsample a single-channel image into multiple multi-channel feature maps of different scales; Extracting low-level features of the multi-channel feature map, wherein the low-level features include at least one of edge, line, corner, etc. level and position information; Based on a normalization algorithm, normalizing the multi-channel feature map; Based on a linear rectification activation function, a nonlinear transformation is performed on the multi-channel feature map, and the transformed multi-channel feature map is output.

3. The infrared small target detection method according to claim 1, characterized in that: The step of extracting the detail features of the multi-channel feature map based on a convolution operation and outputting a first feature map comprises: Reducing the number of channels of the multi-channel feature map using convolution; Reducing the scale of the multi-channel feature map after convolution through a fully connected layer and extracting detail features; Normalizing the multi-channel feature map and restoring the scale of the multi-channel feature map; Restoring the number of channels of the multi-channel feature map using convolution; The multi-channel feature map is normalized, and the first feature map is output.

4. The infrared small target detection method according to claim 1, characterized in that: The step of upsampling the second feature map by using a bilinear interpolation method and obtaining a third feature map includes: Compressing the number of channels of the second feature map; The second feature map after normalized convolution; Based on a linear rectification function, performing a nonlinear transformation on the second characteristic map; The second feature map output by the linear rectification function is upsampled by using a bilinear interpolation method, and a third feature map is obtained.

5. The infrared small target detection method according to claim 1, characterized in that: The step of integrating the multi-channel features of the fused feature map and converting the convolved multi-channel features into binary classification probabilities of pixel points includes: Integrate the multi-channel features of the fused feature map and reduce the dimension of the feature tensor; The fused feature map after normalized convolution; Based on a linear rectification function, performing a nonlinear transformation on the fused feature map; The number of channels of the fused feature map is compressed by convolution to obtain a single-channel feature map; The multi-channel features of the fused feature map are integrated, and the convolved features are converted into binary classification probabilities of pixel points.

6. The infrared small target detection method according to claim 1, characterized in that: Before the step of downsampling a single-channel image into a plurality of multi-channel feature maps of different scales by using convolution, the method further includes: Unify the input images into single-channel images of a preset scale; When the scale of the image is larger than the preset scale, the middle area of ​​the image is cropped. When the scale of the image is smaller than the preset scale, the missing area is filled with a constant value of 0.

7. An electronic device, characterized in that: include: A memory for storing executable program codes; as well as A processor is connected to the memory and runs a computer program corresponding to the executable program code by reading the executable program code to execute the steps in the infrared small target detection method according to any one of claims 1 to 6.

8. An infrared small target detection system, characterized in that: Comprising the electronic device as claimed in claim 7.