CN-121999263-A - Detection method and system for infrared small target

CN121999263ACN 121999263 ACN121999263 ACN 121999263ACN-121999263-A

Abstract

The invention belongs to the field of target detection, and particularly relates to a detection method and a detection system of an infrared small target. The method comprises the steps of 1) inputting images to be detected into a detection model to obtain detection results of infrared small targets, wherein the training mode of the detection model comprises the steps of 1) inputting images in a training set, generating a suggestion frame according to an extracted feature image, 2) respectively obtaining decision values of targets and background classifications of frame selection contents in the suggestion frame according to a decision algorithm corresponding to average contrast and average second-order gradient in priori significance decisions, obtaining decision values of targets and background classifications of frame selection contents according to CNN decisions, carrying out weighted summation on the decision values obtained by the three according to the algorithm, the CNN decisions and weights corresponding to the CNN decisions to obtain decision results, repeating the steps until the decision results of all the suggestion frames are obtained, 3) obtaining detection results according to frame regression prediction results and decision results, updating the model according to the detection results, the suggestion frames and real labels, and iterating 1) -3) until stopping conditions are met.

Inventors

SUN MAOXIANG
JIN YONG
REN JING
MA QI
WU JIAN
CHENG YUAN
Guan Rixin
HE ZHIGUO
XU WEIMING
WANG ZHIFANG
ZHU YI
ZHAO YING

Assignees

华能太仓发电有限责任公司
河南许继电力电子有限公司

Dates

Publication Date: 20260508
Application Date: 20241108

Claims (8)

1. A detection method of an infrared small target is characterized in that an image to be detected is input into a detection model to obtain a detection result of the infrared small target, and a training mode of the detection model comprises the following steps: 1) Inputting images in a training set, and generating a suggestion frame according to a feature image obtained by feature extraction; 2) Obtaining decision values of the target and the background classification of the frame selection content in a single suggestion frame of an input image added with the suggestion frame according to a decision algorithm corresponding to the average contrast in the prior significance decision and a decision algorithm corresponding to the average second-order gradient, obtaining the decision values of the target and the background classification of the frame selection content according to the CNN decision, carrying out weighted summation on the decision values of the target and the background classification obtained by the three according to the weight corresponding to the average contrast algorithm, the decision corresponding to the average second-order gradient algorithm and the CNN decision, obtaining the decision results of the target and the background classification of the frame selection content in the suggestion frame, and repeating the steps until the decision results of all the suggestion frames are obtained; 3) Obtaining a detection result of the infrared small target according to a result of carrying out frame regression prediction on the suggestion frame and the decision result; and (3) updating a detection model according to the detection result of the infrared small target, the generated suggestion frame and the real label, and iterating 1) -3) until the stopping condition is met.
2. The method for detecting the infrared small target according to claim 1, wherein the mode of the feature map obtained by feature extraction comprises the steps of carrying out N layers of downsampling on an input image through a backbone network, carrying out weighted fusion on features obtained by upsampling on the N layers and features before downsampling on the N layers, carrying out weighted fusion on a weighted fusion result, carrying out weighted fusion on the obtained result and the features before downsampling on the N-1 layers, and repeatedly carrying out calculation of carrying out weighted fusion on the latest weighted fusion result, carrying out weighted fusion on the latest weighted fusion result and the features before downsampling on the previous layer until the latest weighted fusion result is subjected to weighted fusion, and carrying out weighted fusion on the latest weighted fusion result and the features before downsampling on the 2 layers, thereby obtaining the feature map.
3. The method for detecting the small infrared target according to claim 1 or 2, wherein the method for obtaining the decision value of the target and the background classification of the content in the suggestion frame according to the CNN decision comprises the steps of classifying the target and the background classification of the content in the suggestion frame according to the CNN decision, and obtaining the decision value of the target and the background classification of the content in the suggestion frame.
4. The method for detecting an infrared small target according to claim 1 or 2, wherein the method for updating the detection model according to the detection result of the infrared small target, the generated suggestion frame and the real tag comprises: And updating the detection model by taking the total loss function which contains a loss function item corresponding to the difference between the detection result of the infrared small target and the real label and a loss function item corresponding to the difference between the generated suggestion frame and the real label as targets.
5. The method for detecting an infrared small target according to claim 1 or 2, wherein the result of performing frame regression prediction on the suggestion frame is obtained by performing frame regression prediction on the suggestion frame by using the CNN decision.
6. The method for detecting the small infrared target according to claim 1 or 2, wherein the method for generating the suggestion frame according to the feature map obtained by feature extraction comprises the steps of inputting the feature map obtained by feature extraction into an RPN network to obtain the generated suggestion frame.
7. The method for detecting small infrared targets according to claim 2, wherein ConvNeXt blocks are used to construct the backbone network.
8. A detection system for small infrared targets comprising a processor, characterized in that the processor is adapted to execute a computer program for carrying out the steps of the detection method for small infrared targets according to any one of claims 1-7.

Description

Detection method and system for infrared small target Technical Field The invention belongs to the field of target detection, and particularly relates to a detection method and a detection system of an infrared small target. Background In recent years, with social progress, the demand for high-performance small-target detection technology in the military, industrial, and civil fields has increased. In particular, in the night or severe weather, the conventional visible light detection is limited, and the infrared small target (fig. 1 is a sample illustration of the infrared small target) detection technology can provide stable and continuous detection capability, and shows important application value. However, when in remote detection, the problems of few target pixels, fuzzy characteristics, easiness in clutter and heat source interference and the like are solved, and a great challenge is brought to the detection of the infrared small target. In the past decades, technicians have also proposed different technical means for detecting small infrared targets, and the underlying logic of these technical means is to design feature extractors and detectors by understanding the data structure and empirically set super parameters. However, the manner of setting the super-parameters according to experience often results in inaccurate and reliable setting of the super-parameters due to the problems of high limitation of artificial factors, high complexity of the system, insufficient adaptability, flexibility, accuracy and reliability, etc. Meanwhile, the methods depend on surface features such as gray values, lack of recognition of deep semantic differences between the target and the background, and therefore poor effect of detecting the fuzzy target under a complex background is achieved. In recent years, deep learning, particularly Convolutional Neural Networks (CNNs), has achieved significant success in target detection. Compared with the traditional mode, the CNN-based target detection method can integrate feature extraction, fusion and classification to realize end-to-end detection, and can acquire deep semantic information from a complex image so as to improve detection accuracy, so that a Convolutional Neural Network (CNN) is applied to infrared small target detection, but a series of problems still exist in the technology in practice. The most troublesome problem of applying Convolutional Neural Network (CNN) to infrared small target detection is that the characteristics of the infrared small target are weak. These weak small objects tend to appear in the form of low contrast, low signal-to-noise ratio in the infrared image, with blurred edges, no prominent shape details, and relatively small dimensions. The faintness in such features makes them extremely prone to masking or aliasing in complex background environments, especially where the background is subject to strong interference (e.g., clutter, thermal noise, dynamic changes, etc.), the signal strength of small objects is more difficult to distinguish from the background. Disclosure of Invention The invention aims to provide a detection method and a detection system for an infrared small target, which are used for solving the problems of larger detection error and poorer detection effect of a detection model caused by the fact that the characteristics of the infrared small target are weak and are easily covered or confused in a complex background environment in the prior art. In order to achieve the above purpose, the invention provides a method for detecting an infrared small target, which specifically comprises the steps of inputting an image to be detected into a detection model to obtain a detection result of the infrared small target, wherein the training mode of the detection model comprises the following steps: 1) Inputting images in a training set, and generating a suggestion frame according to a feature image obtained by feature extraction; 2) Obtaining decision values of the target and the background classification of the frame selection content in a single suggestion frame of an input image added with the suggestion frame according to a decision algorithm corresponding to the average contrast in the prior significance decision and a decision algorithm corresponding to the average second-order gradient, obtaining the decision values of the target and the background classification of the frame selection content according to the CNN decision, carrying out weighted summation on the decision values of the target and the background classification obtained by the three according to the weight corresponding to the average contrast algorithm, the decision corresponding to the average second-order gradient algorithm and the CNN decision, obtaining the decision results of the target and the background classification of the frame selection content in the suggestion frame, and repeating the steps until the decision results of all the suggestion frames are