CN-122023783-A - Method, device, equipment and medium for detecting infrared small target

CN122023783ACN 122023783 ACN122023783 ACN 122023783ACN-122023783-A

Abstract

The application provides a detection method, device, equipment and medium for infrared small targets, which are characterized in that after an infrared image of a target environment is acquired, a target detection optimization model is constructed based on an initial detection optimization model and image content information of the infrared image, the target detection optimization model comprises a plurality of optimization variables, the plurality of optimization variables comprise main optimization variables used for representing the image and auxiliary optimization variables, the auxiliary optimization variables are used for updating the auxiliary main optimization variables, the main optimization variables comprise image basic variables and first low-rank decomposition variables, the plurality of optimization variables are iteratively updated by using an alternating direction multiplier method until preset iteration requirements are met, target values of the image basic variables are obtained, whether the small targets exist in the target environment is determined, the image basic variables are updated based on the auxiliary optimization variables, and the auxiliary optimization variables are updated based on the first low-rank decomposition variables by using a matrix broadcasting mechanism. According to the embodiment of the application, the detection efficiency of the infrared small target can be improved.

Inventors

ZHANG CHAOZI
LI CHAO
LI LU
CHEN YUE

Assignees

之江实验室

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (11)

1. The method for detecting the infrared small target is characterized by comprising the following steps of: Acquiring an infrared image of a target environment, and inputting the infrared image into a target neural network, wherein the target neural network comprises an image processing network and an image detection network, and the image detection network is integrated with an initial detection optimization model; Performing image information extraction processing on the infrared image by using the image processing network to obtain image content information; constructing a target detection optimization model for the infrared image based on the initial detection optimization model and the image content information, wherein the target detection optimization model comprises a plurality of optimization variables, the plurality of optimization variables comprise main optimization variables used for representing the image and auxiliary optimization variables used for assisting the updating of the main optimization variables, and the main optimization variables comprise image basic variables and first low-rank decomposition variables; And in each iteration period, updating the image basic variable based on the auxiliary optimization variable, updating the auxiliary optimization variable based on the first low-rank decomposition variable by using a matrix broadcasting mechanism, and determining whether a small target exists in the target environment by the target value of the image basic variable.
2. The method of claim 1, wherein the main optimization variables further comprise a second low-rank decomposition variable, the image base variables comprise image coefficient variables, noise variables and target variables, the iteratively updating the plurality of optimization variables using an alternating direction multiplier until a preset iteration requirement is met, the obtaining the target values of the image base variables comprises: updating, with a solver, the first low-rank decomposition variable based on a second low-rank decomposition variable value and an image coefficient variable value of a current iteration period; updating the auxiliary optimization variable based on the first low rank decomposition variable using a matrix broadcasting mechanism; Based on the updated auxiliary optimization variables, respectively updating the image coefficient variables, the noise variables and the target variables to obtain updated values of the image coefficient variables, the noise variables and the target variables; Repeatedly executing the steps until the preset iteration requirement is met under the condition that the updated value of the image basic variable does not meet the preset iteration requirement, and obtaining the target value of the image coefficient variable, the target value of the noise variable and the target value of the target variable; a target value of the image base variable is determined based on the target value of the image coefficient variable, the target value of the noise variable, and the target value of the target variable.
3. The method of claim 2, wherein the updating auxiliary optimization variables based on the first low rank decomposition variables using a matrix broadcast mechanism comprises: Based on the first low-rank decomposition variable, respectively constructing a first auxiliary matrix, a second auxiliary matrix and a third auxiliary matrix by adopting a tensor broadcasting mechanism; And determining auxiliary optimization parameters based on the first auxiliary matrix, the second auxiliary matrix and the third auxiliary matrix, and updating the auxiliary optimization variables based on the auxiliary optimization parameters by using a soft threshold method.
4. A method according to claim 3, wherein each row of the first auxiliary matrix is repeatedly filled in the same order based on the modulo-long square of each column vector of the first low-rank decomposition variable, each column of the second auxiliary matrix is repeatedly filled in the same order based on the modulo-long square of each column vector of the first low-rank decomposition variable, and the third auxiliary matrix is determined based on the inner product of the first low-rank decomposition variable.
5. The method of any one of claims 1-4, wherein after the obtaining the image base variable, the method further comprises: And carrying out image reconstruction based on the image basic variables to obtain a characteristic image, carrying out threshold segmentation on pixels in the characteristic image to obtain a mask image, and determining whether a small target exists in the target environment according to the mask image.
6. The method of claim 5, wherein the method further comprises: And under the condition that a small target exists in the target environment, determining the position of the small target in the target environment, and monitoring whether the small target enters a preset area range according to the position.
7. The method according to claim 1, wherein the image information extraction processing is performed on the infrared image by using the image processing network to obtain image content information, including: Performing sliding segmentation on the infrared image to obtain a plurality of block images, converting each block image into column vectors, and constructing block image matrix information corresponding to the infrared image based on the column vectors corresponding to the block images respectively; Extracting edge information of the infrared image to obtain overlapped edge information; And obtaining the image content information based on the block image matrix information and the overlapped edge information.
8. The method of claim 7, wherein the image base variables include a background variable, a noise variable, and a target variable, wherein the constructing a target detection optimization model for the infrared image based on the initial detection optimization model and the image content information comprises: substituting the background component, the noise component and the target component in the block image matrix information into the background variable, the noise variable and the target variable in the initial detection optimization model, and taking the edge information as the weight of the target variable to obtain the target detection optimization model.
9. A detection device for small infrared targets, comprising: the image acquisition module is used for acquiring an infrared image of a target environment and inputting the infrared image into a target neural network, wherein the target neural network comprises an image processing network and an image detection network, and the image detection network is integrated with an initial detection optimization model; the image processing module is used for extracting and processing the image information of the infrared image by utilizing the image processing network to obtain image content information; A model construction module, configured to construct a target detection optimization model for the infrared image based on the initial detection optimization model and the image content information, where the target detection optimization model includes a plurality of optimization variables, the plurality of optimization variables include a main optimization variable for characterizing an image and an auxiliary optimization variable for assisting the main optimization variable in updating, and the main optimization variable includes an image base variable and a first low-rank decomposition variable; And in each iteration period, the image basic variable is updated based on the auxiliary optimized variable, the auxiliary optimized variable is updated based on the first low-rank decomposition variable by using a matrix broadcasting mechanism, and the target value of the image basic variable is used for determining whether a small target exists in the target environment.
10. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the steps of the method for detecting small infrared objects according to any one of claims 1-8 when said program is executed by said processor.
11. A computer readable storage medium having stored thereon a computer program, characterized in that the program when executed by a processor realizes the steps of the method for detecting small infrared targets according to any of claims 1-8.

Description

Method, device, equipment and medium for detecting infrared small target Technical Field The application relates to the technical field of computer vision, in particular to a method, a device, equipment and a storage medium for detecting an infrared small target. Background An infrared search and tracking system (IRST) is widely applied to the fields of unmanned aerial vehicle navigation, space detection and the like. However, small infrared targets are usually characterized by small size, low brightness, and complex background clutter, and it is difficult for conventional filtering or detection methods based on visual saliency to effectively distinguish targets from backgrounds. In the related art, the Self-regularized weighted sparse Model (Self-Regularized WEIGHTED SPARSE Model, SRWS) is based on a multi-subspace hypothesis, and the Self-regularized term and the weighted sparse term are introduced to realize the robust target extraction under a complex background, however, a large number of matrix singular value decomposition (Singular Value Decomposition, SVD) is often required in the calculation process in the mode, the calculation complexity is high, and the image detection efficiency is low. Disclosure of Invention In view of the above, the present application provides a method, apparatus, device and storage medium for detecting small infrared targets, so as to at least solve the problems in the related art. Specifically, the application is realized by the following technical scheme: the application provides a detection method of an infrared small target, which comprises the following steps: Acquiring an infrared image of a target environment, and inputting the infrared image into a target neural network, wherein the target neural network comprises an image processing network and an image detection network, and the image detection network is integrated with an initial detection optimization model; Performing image information extraction processing on the infrared image by using the image processing network to obtain image content information; constructing a target detection optimization model for the infrared image based on the initial detection optimization model and the image content information, wherein the target detection optimization model comprises a plurality of optimization variables, the plurality of optimization variables comprise main optimization variables used for representing the image and auxiliary optimization variables used for assisting the updating of the main optimization variables, and the main optimization variables comprise image basic variables and first low-rank decomposition variables; And in each iteration period, updating the image basic variable based on the auxiliary optimization variable, updating the auxiliary optimization variable based on the first low-rank decomposition variable by using a matrix broadcasting mechanism, and determining whether a small target exists in the target environment by the target value of the image basic variable. In some embodiments, the main optimization variable further includes a second low-rank decomposition variable, the image base variable includes an image coefficient variable, a noise variable, and a target variable, and the iteratively updating the plurality of optimization variables by using the alternate direction multiplier method until a preset iteration requirement is satisfied, to obtain a target value of the image base variable includes: updating, with a solver, the first low-rank decomposition variable based on a second low-rank decomposition variable value and an image coefficient variable value of a current iteration period; updating the auxiliary optimization variable based on the first low rank decomposition variable using a matrix broadcasting mechanism; Based on the updated auxiliary optimization variables, respectively updating the image coefficient variables, the noise variables and the target variables to obtain updated values of the image coefficient variables, the noise variables and the target variables; Repeatedly executing the steps until the preset iteration requirement is met under the condition that the updated value of the image basic variable does not meet the preset iteration requirement, and obtaining the target value of the image coefficient variable, the target value of the noise variable and the target value of the target variable; a target value of the image base variable is determined based on the target value of the image coefficient variable, the target value of the noise variable, and the target value of the target variable. In some implementations, the updating auxiliary optimization variables based on the first low rank decomposition variable using a matrix broadcast mechanism includes: Based on the first low-rank decomposition variable, respectively constructing a first auxiliary matrix, a second auxiliary matrix and a third auxiliary matrix by adopting a tensor broadcasting mechanism; And d