Deep Learning Based Infrared Small Target Detection Algorithm Research

Infrared small target detection technology is widely used in the military,industry,agriculture,transportation and other fields,infrared imaging with high accuracy,high concealment,clear imaging,strong anti-interference ability,etc.,and therefore has significant research value,in recent years thanks to the development of strong algorithm theory and feature analysis technology,represented by YOLO and Faster R-CNN,deep learning-based detection technology has gained in recent years,thanks to the development of strong algorithm theory and feature analysis techniques,deep learningbased detection techniques,represented by YOLO and Faster R-CNN,have made great progress.However,these methods only acquire features by traversing local images,resulting in low detection accuracy,limited detection types and processing speed to meet real-time requirements,which is still a long way from the requirements of practical applications.In order to improve the detection effect of infrared small target detection algorithm in different scenes,two improved algorithms are proposed in this thesis,and the main research results are as follows:(1)The detection framework mainly consists of a generative adversarial network and a deep unfolding network,and the designed network generator partly consists of a deep unfolding network,so it is called a deep unfolding adversarial network.For the input infrared images,they enter two generators,which are composed of deep unfolding networks and can effectively target the false alarm and missed detection problems by controlling the number of unfolding layers,respectively.The discriminator,on the other hand,plays a supervisory role and can effectively optimize the network to improve the overall detection effect.(2)In this thesis,we also propose an adaptive dictionary-based multi-scale coded-decoded infrared small target detection network to address the current challenges of infrared small target detection technology,whose overall structure consists of a coding and decoding network and an adaptive dictionary transform module.In the detection process,the infrared images will first enter the coding network,which will output different scales of infrared images and feature maps to the decoding network,where the feature maps are processed by the adaptive dictionary transform module.The decoder then fuses the acquired inputs and finally outputs the detection results after processing.In order to study the performance and effectiveness of the algorithm proposed in this thesis,the data set in this thesis uses real infrared background images and covers a wide range of images with different weather conditions and different shooting ranges,and the effectiveness of the infrared small target detection algorithm proposed in this thesis is verified through subjective and objective experimental comparison and analysis.