Research On Infrared Dim And Small Target Detection Method Based On Concolutional Neural Network

Infrared dim and small target detection has an important role in many fields.The task faces a challenging topic with small target scale,weak detected signal,low signalto-noise ratio easily drowned by the background,few features without specific shape,inability to obtain its texture information,and the existence of a large number of structural noise interference phenomena in the background.This paper addresses the difficulties of infrared dim and small target detection by using convolutional neural network as the main detection method,and the following research work is carried out to study the subject from two parts: single-frame detection and multi-frame association,respectively.Firstly,aiming at the problem of high difficulty and few samples of infrared dim and target detection data,a target simulation synthesis method is designed.And combined with imaging characteristics,fusion of defocus blur,motion blur and other target data enhancement methods,closer to the real target but also expand the sample set.Secondly,the infrared dim and small target detection method for single frame image is studied.The convolution neural network is used as the main framework of the detection process,and the classical network module structure is introduced.The network architecture suitable for infrared dim and small target detection task is designed.The detection task is modeled as pixel-level segmentation task and key point detection task.The corresponding loss function is used to train the two tasks,and the segmentation level output detection algorithm and the instance level output detection algorithm are used for subsequent verification.Then,according to different output representations and labeling methods in the single frame detection of infrared dim and small targets,ROC curve is used to measure the performance.Three calculation methods of ROC curve are designed,including segmentation-level labeling of segmentation-level algorithm,instance-level labeling of segmentation-level algorithm,and instance-level labeling of instance-level algorithm.According to the calculation principle of ROC curve,the three calculation methods are optimized to accelerate the evaluation calculation efficiency.Finally,aiming at the problem of infrared dim and small target detection in sequence images,based on the form of multi-frame association,aiming at the contradiction between detection rate and false alarm rate,a cascade detection algorithm based on multi-frame association is proposed.The detection algorithm is divided into high priority detection algorithm and low priority detection algorithm,which are responsible for low false alarm rate and high detection rate.The detection process based on different standards is realized.