CN-121982528-A - Lightweight infrared small target detection enhancement method, device and medium integrating local contrast and edge information

CN121982528ACN 121982528 ACN121982528 ACN 121982528ACN-121982528-A

Abstract

The invention discloses a light infrared small target detection enhancement method, device and medium integrating local contrast and edge information in the technical field of artificial intelligence, wherein the method comprises the steps of obtaining an infrared image to be processed; and inputting the infrared image to be processed into a trained infrared small target detection network to obtain a prediction label of a small target on the infrared image. The light infrared small target detection enhancement method, device and medium for fusing the local contrast and the edge information can greatly improve the extraction and retention capacity of the infrared small target detection backbone network to the infrared small target position information and the structural characteristics, realize the decoupling of the processing module of the local contrast priori and the edge priori and the backbone network, and effectively enhance the universality and the flexibility of the method.

Inventors

ZHANG RUI
WANG PEICHAO
WANG JIABAO
LI YANG
MIAO ZHUANG

Assignees

中国人民解放军陆军工程大学

Dates

Publication Date: 20260505
Application Date: 20260121

Claims (10)

1. A light infrared small target detection enhancement method integrating local contrast and edge information is characterized by comprising the following steps: acquiring an infrared image to be processed; Inputting the infrared image to be processed into a trained infrared small target detection network to obtain a prediction label of a small target on the infrared image; the processing of the infrared image to be processed by the trained infrared small target detection network comprises the following steps: Reconstructing the infrared image to be processed into a single-channel strengthening characteristic diagram; Extracting global features from the single-channel enhanced feature map; extracting edge weight of an infrared image to be processed; Enhancing the global feature according to the edge weight; And generating a detection mask according to the enhanced global features to obtain a final prediction tag.
2. The method for enhancing light-weight infrared small target detection by fusing local contrast and edge information according to claim 1, wherein the method is characterized in that before the infrared image to be processed is input into a trained infrared small target detection network, the infrared image to be processed is further preprocessed, and the preprocessing comprises the step of normalizing the infrared image to be processed.
3. The method for enhancing light-weighted infrared small target detection by fusing local contrast and edge information according to claim 1, wherein reconstructing the infrared image to be processed into a single-channel enhanced feature map comprises: extracting a local contrast attention weight matrix from an infrared image to be processed; and splicing the local contrast attention weight matrix with the infrared image to be processed in the channel dimension, and integrating and enhancing to obtain a single-channel enhanced feature map.
4. The method for enhancing light-weighted infrared small target detection by fusing local contrast and edge information as set forth in claim 3, wherein the extracting the local contrast attention weight matrix from the infrared image to be processed comprises: calculating an infrared image to be processed Upper coordinates are The calculation formula of the local contrast distance value of the points of (a) is as follows: ; Wherein, the The number of channels representing the infrared image is 1, the height and width are H and W respectively, A local contrast distance value representing a first direction, A local contrast distance value representing a second direction, A local contrast distance value representing a third direction, The local contrast distance value in the fourth direction is represented, wherein the first direction is the left-upper-right downward direction, the second direction is the right-upper-left downward direction, the third direction is the up-lower direction, the fourth direction is the left-right direction, Representation of The abscissa of the upper point, and , Representative of The ordinate of the upper point, and , The number of pixels from the center is taken in the convolution, And Super parameters which are all larger than 0; Calculation of Upper coordinates are Local contrast attention weight value of points of (2) Obtaining Is a local contrast attention weight matrix of (2) , The calculation formula of (2) is as follows: 。
5. The method for detecting and enhancing the light-weighted infrared small target by fusing local contrast and edge information according to claim 4, wherein the local contrast attention weight matrix and the infrared image to be processed are spliced in the channel dimension, and are integrated and enhanced to obtain a single-channel enhanced feature map, and the calculation formula is as follows: ; Wherein, the Representing a single-channel enhanced feature map, CBP represents a sequential combination of a3 x 3 convolutional layer, a BN layer, and a PReLU activation function, Is a local contrast attention weight matrix And input image The resulting fused feature map is superimposed in the channel dimension, Representing a 3 x 3 convolution of the data, Representing a convolution operation.
6. The method for lightweight infrared small target detection enhancement by combining local contrast and edge information according to claim 5, wherein extracting edge weights of an infrared image to be processed, enhancing global features according to the edge weights, comprises: Extracting target edge information from an infrared image to be processed to obtain an edge weight feature map; the edge weight feature map is multiplied element by element with the global feature map.
7. The method for enhancing light-weighted infrared small target detection by combining local contrast and edge information as set forth in claim 6, wherein the step of extracting target edge information from the infrared image to be processed to obtain an edge weight feature map comprises the steps of utilizing a Sobel operator to obtain the infrared image to be processed Extracting coarse edges to obtain primary edge feature images The calculation formula is as follows: Wherein, the Representing a Sobel operator; Primary edge feature map by multiple cascaded nonlinear transformation operations Extracting and optimizing into high-level edge feature image For a pair of Performing dimension reduction of channel dimension to generate an unnormalized edge weight characteristic diagram The calculation formula is as follows: Activating a function pair by Sigmoid Normalization processing is carried out to obtain a final edge weight feature map The calculation formula is as follows:
8. the method of claim 6, wherein the infrared small target detection network comprises a main branch and an edge branch in parallel, and the main branch comprises: The local contrast attention fusion module is used for reconstructing an input image into a single-channel enhanced feature image and outputting the single-channel enhanced feature image to the global feature extraction module; the global feature extraction module is used for extracting global features from the input enhanced feature graphs and transmitting the global features to the feature fusion module; The feature fusion module is used for enhancing the global features according to the edge weights and transmitting the enhanced global features to the output convolution module; The output convolution module is used for generating a prediction tag according to the enhanced global features; the edge branch includes: The self-adaptive edge weight extraction module is used for extracting the edge weight of the infrared image to be processed and transmitting the edge weight to the feature fusion module.
9. A computer apparatus, comprising: A memory for storing computer programs/instructions; a processor for executing the computer program/instructions to implement the steps of the method of any one of claims 1-8.
10. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the steps of the method according to any one of claims 1-8.

Description

Lightweight infrared small target detection enhancement method, device and medium integrating local contrast and edge information Technical Field The invention relates to the technical field of artificial intelligence, in particular to a light infrared small target detection enhancement method, device and medium for fusing local contrast and edge information. Background In modern war, mastering battlefield situation awareness is a key to acquiring 'control information right' and winning war initiative. The infrared imaging technology plays a vital role in the fields of early warning detection, accurate guidance, reconnaissance monitoring and the like by virtue of the unique advantages of being passive in operation, good in concealment, strong in electromagnetic interference resistance, capable of working all day long and the like. On the battlefield, the infrared small target is often an early manifestation of an enemy attack missile or small unmanned plane threat. Therefore, by means of rapid, accurate and stable detection of the infrared small targets, precious early warning time can be provided for the defense system of the user, the fight efficiency of the weapon system is improved, and the method is a core link for guaranteeing the survivability of the user and has very important application value and strategic significance. In a modern battlefield environment with more and more complexity, how to quickly find small targets with low intensity, blurred edges and less information on an image by using infrared detection equipment is a task with great research value. The existing infrared small target detection method is mainly divided into two main types of single-frame detection and multi-frame detection, and the single-frame detection algorithm has natural robustness to the relative motion of a target and a background because the single-frame detection algorithm does not depend on inter-frame information. This feature enables a stable and reliable initial input for subsequent multi-frame timing analysis, and therefore, single-frame detection methods have been widely and intensively studied. Existing single frame infrared small target detection methods can be broadly divided into two major categories, model-driven (model-driven) methods and data-driven (data-driven) methods. The model driven method generally follows a specific assumption, uses expert knowledge to directly construct a detection model, does not need support of a large amount of data, is greatly influenced by complex background and noise, and has lower robustness. The data-driven method automatically excavates deep semantic information of the target from mass data through end-to-end feature learning by means of a deep learning technology, and shows superior performance far superior to the traditional method in aspects of target feature extraction and complex background suppression. However, the pure data driven paradigm has its own limitations. On one hand, the 'black box' characteristic of the deep learning model enables the deep learning model to neglect the priori knowledge with clear physical meaning of the known infrared small target to a certain extent, and on the other hand, the high dependence of the model on large-scale labeling data also makes the deep learning model challenging in practical application scenes with scarce samples. Disclosure of Invention The invention aims to overcome the defects in the prior art, and provides a light infrared small target detection enhancement method, device and medium combining local contrast and edge information, which can greatly improve the extraction and retention capacity of an infrared small target detection backbone network to infrared small target position information and structural characteristics, realize the decoupling of a processing module of local contrast priori and edge priori and the backbone network, and effectively enhance the universality and flexibility of the method. In order to achieve the above purpose, the invention is realized by adopting the following technical scheme: In a first aspect, the present invention provides a light-weight infrared small target detection enhancement method for fusing local contrast and edge information, including: acquiring an infrared image to be processed; Inputting the infrared image to be processed into a trained infrared small target detection network to obtain a prediction label of a small target on the infrared image; the processing of the infrared image to be processed by the trained infrared small target detection network comprises the following steps: Reconstructing the infrared image to be processed into a single-channel strengthening characteristic diagram; Extracting global features from the single-channel enhanced feature map; extracting edge weight of an infrared image to be processed; Enhancing the global feature according to the edge weight; And generating a detection mask according to the enhanced global features to obtain a