Research On Infrared Small Target Detection Technology Based On Local Gray Features

The infrared point target image is collected by the Infrared Search and Tracking System(IRST).The infrared search and tracking system is mainly used to achieve the functions of target search,detection and tracking,to find and track targets as far as possible,providing a longer response time for further attacks.Due to the excellent performance of infrared small target detection technology in the fields of guidance and early warning systems,the research of infrared small target detection and its application in the military is more meaningful.Due to the long imaging distance,the infrared small target occupies only a small number of pixels on the infrared image,and the target is dot-like,textureless,and has a low contrast with the background.In recent years,scholars at home and abroad have put forward a large number of infrared small target detection algorithms for this problem,and have achieved good results.However,the infrared small target data set used by most methods lacks diversity,that is,the background of the infrared image is mostly a single scene,which leads to the low robustness of the algorithm.It is precisely because of the lack of infrared small target data sets containing multiple scenes that the research team manually annotates infrared small targets in infrared images.In order to solve the problem of infrared small target detection difficulty in complex background,this paper puts forward the infrared small target detection algorithm based on local covariance matrix and the infrared small target detection algorithm based on the improved Xception model,which are optimized in detection performance.The details of the study are as follows:(1)Because the infrared small target is dot-like in the infrared image,occupies fewer pixels,and there is a lot of background noise and system noise in the image,the image is first filtered by Gaussian in order to reduce noise interference to the algorithm.Then,the gradient change of the local area of the image pixel small is calculated from multiple directions,the local covariance matrix is recorded,and the size and direction of the gradient change are calculated.According to the size of gradient change and the direction of gradient change to determine whether it is a suspected target,and finally use local mean and variance for adaptive threshold segmentation to remove pseudo-target,to obtain the location of infrared small target.Experimental results show that the performance of infrared small target detection algorithm based on local covariance matrix is better than that of comparison experimental algorithm.(2)Infrared image system in the imaging process there are a lot of background interference,greatly increased the difficulty of infrared small target detection and detection algorithm of the false alarm rate is high.In order to solve this problem,this paper proposes to detect infrared small targets based on the neural network structure of the improved Xception structure.Because of the small number of infrared small target data sets,the Image Net dataset is used for pretraining and the pretrained model is obtained.The infrared small target dataset is then fine-tuned on the pretrained model,and finally the model outputs the location of the infrared small target.In this paper,the improvement of network performance by data augmentation is analyzed,and the experiment is compared with the mainstream network model.Experimental results show that the results based on improving Xception's infrared small target detection algorithm are more effective than those based on local covariance matrix,and the detection performance is better than that of YOLO series algorithms.