Infrared Small Target Detection Based On Multi-scale Attentional Feature Fusion

It is the fundamental task of the infrared detection system to extract the target features according to the real-time infrared image,and to detect and identify the target as soon as possible.The performance of infrared target detection technology is mainly reflected in the detection performance of small targets.There are many technical difficulties in infrared small target detection,such as too small target size,scarce intrinsic features,complex imaging background,low image contrast and signal-to-noise ratio,etc.In view of the above problems,this thesis further explores how to improve the comprehensive performance of infrared small target detection from the two directions of model-driven method and data-driven method,in order to achieve robust detection in various complex scenes.Detailed contributions are summarized as follows:(1)The imaging principle and characteristics of infrared dim and small targets are studied.It mainly includes infrared target imaging theory,infrared dim and small target characteristics analysis and related data set research,which lays the foundation for the subsequent method design of this thesis;(2)Infrared small object detection method based on multi-attentional local contrast is investigated.Firstly,the working process of the human visual system and the visual attention mechanism,contrast mechanism,field of view pop-up mechanism and other visual characteristics are expounded from the perspectives of physiology and psychology.Inspired by the characteristics of human vision,this algorithm further mines the characteristics of small infrared targets and their background neighborhoods,and constructs more discriminative contrast metrics from different perspectives,such as weighted feature fusion and direction selectivity.Finally,the algorithm is compared with other algorithms,which proves that it has better performance in target enhancement,background suppression and detection accuracy;(3)Infrared small object detection based on local contrast prior to guide regional attention is studied.Starting from the classic deep learning target detection algorithm and feature fusion framework,this thesis analyzes the technical difficulties of the current target detection network when it is used for small infrared target detection tasks.According to the above problems,a multi-scale coordinate attention feature fusion module is proposed,which retains the detailed information of the region of interest in the deep network,and alleviates the problem of target loss in the deep network to a certain extent.In addition,a candidate region generation network guided by local contrast prior is designed,which can screen out a large number of weakly significant regions and save computing resources under the premise of ensuring the recall rate.Finally,through various experiments,it is verified that the algorithm is better than the algorithm that only uses model-driven or data-driven in various scenarios,which reflects the effectiveness of integrating model priors into data-driven algorithms.