Research On Infrared Small Target Detection Technology Based On Deep Learning

Infrared small target detection is a key technology for environment awareness on unmanned platforms,and is of great importance in military fields such as infrared search and tracking,navigation and guidance,and intelligence reconnaissance.The typical application scenario requirements and infrared imaging characteristics make the target in the infrared image has the characteristics of few pixels,weak texture information,low signal-to-noise ratio and serious background clutter interference,etc.Small target detection faces great difficulties and challenges in terms of accuracy and robustness.Aiming at the important and difficult problems of single-frame infrared small target detection,this paper thoroughly researches the deep learning-based infrared small target detection algorithm,which effectively improves the detection capability of the unmanned vehicle platform vision detection system for infrared small targets.The main research content is as follows:(1)The research background and significance of infrared small target detection are summarized,the characteristics of typical infrared small target images are analyzed,and the important and difficult problems faced by infrared small target detection are condensed.In addition,the basic theoretical knowledge based on deep learning is introduced,and two typical deep learning algorithms for infrared small target detection and the evaluation indexes of infrared small target detection algorithms in this paper are elaborated.(2)A feature pyramid network based on feature fusion is designed for problems such as weak texture information and few available features of infrared small targets,,weak robustness and low detection accuracy of traditional infrared small target detection methods to scene changes.The feature pyramid network is used to extract the features of different layers of the infrared small target,which are fused by means of shuffle connections the shallow and deep features of infrared small targets are fully utilized.Compared with traditional methods,the detection accuracy of infrared small targets is significantly improved.(3)To address the problems of redundant information in the feature fusion module and easy loss of small target features in the deep network,an asymmetric attention feature fusion module is designed and embedded into the U-Net network,effectively overcoming the contradiction between feature resolution and network hierarchy and realizing cross-layer exchange of high-level semantic information and target detail information,as well as channel and spatial information exchange between features of the same layer.The bottom-up global channel attention module achieves dynamic weighted modulation between high-level semantic features and low-level detail features,and the shuffle attention unit achieves information exchange between channel features and spatial features at the same layer,thus effectively preserving infrared small target features and avoiding infrared small targets being overwhelmed by complex background noise.The method achieves the best detection results in terms of Io U,n Io U and PR curves.(4)An unmanned vehicle platform vision detection system for typical tasks is built,and a deep learning-based infrared small target detection algorithm is deployed and debugged on the on-board NVIDIA Xavier embedded image processing platform.The algorithm proposed in this paper achieves a detection speed of 0.04 seconds for an image resolution of 512×512 on the NVIDIA Xavier embedded platform,and the accuracy and speed of the algorithm detection can meet the basic task requirements of the unmanned vehicle.