Research On Infrared Small Target Detection Methods Based On Local And Global Feature Representation

The infrared search and tracking(Infrared Search and Tracking,IRST)system has been widely applied in airspace surveillance,maritime surveillance,aerial early warning,and missile defence due to its advantages of concealment,anti-interference,and all-day work.As one of the critical technologies of the IRST system,the task of infrared small target detection is to detect targets from long-distance infrared imaging,and the result will directly affect the performance of the IRST system.Therefore,the research of infrared small target detection technology is of great significance.In practical scenarios,the research on infrared small target detection faces the difficulties of small target size,scarce texture information,complicated image background and low image signal-to-noise ratio(SNR).With excellent target enhancement and background suppression capabilities,infrared small target detection algorithms based on feature representation have received more and more attention.However,constructing effective feature representations that describe small targets' inherent characteristics appeals to be the critical challenge of designing robust detection algorithms.This thesis focuses on the research of infrared small target detection algorithms based on local and global feature representation because of the above problems.The main research contents of this thesis are summarized as follows:(1)Aiming at the problem that the traditional feature representation based on the assumption of local brightness differences is challenging to enhance small targets in multiple dimensions,this thesis proposes a small target detection algorithm based on multi-scale local brightness difference measure and local energy factor.By introducing the concept of regional local energy,the local energy factor measures the dissimilarity between the target area's structure and its neighbouring areas' structure.The local energy factor is then fused with the local brightness difference measure at multiple scales,which can effectively enhance the target signal.Experimental results show that compared with the existing algorithms that only consider local brightness characteristics,this algorithm shows better target enhancement ability and obtains higher detection accuracy.(2)Aiming at the problem that the existing algorithms based on local feature representation may mistakenly detect strong clutters with similar local characteristics to small targets,this thesis proposes a two-stage small target detection algorithm based on an improved random walk model.At the first stage,we select the target candidates from the whole image based on local features.At the second stage,we detect the targets with local structure showing global rarity based on the improved random walk model.Experimental results show that this algorithm can eliminate more false positives and reduce the false alarm rate compared with the algorithms based on local feature representation.(3)Aiming at the problem that the two-stage algorithm faces the detection bottleneck in the first stage,this thesis proposes an end-to-end small target detection algorithm based on graph Laplacian model with local feature constraints.The algorithm establishes a graph Laplacian model for the input infrared image and outputs the feature representation that describes small targets' global characteristics.Besides,two local feature descriptors are introduced to the model as optimization function constraints,making the model available for describing both local and global characteristics of small targets.Experimental results show that this algorithm obtains higher detection accuracy compared with the existing algorithms.(4)Aiming at the problem that small target detection algorithms fail easily to detect weak targets in multi-target detection scenarios,this thesis proposes a multi-target detection algorithm based on a hierarchical maximal entropy random walk model.We demonstrate the feasibility of the maximal entropy random walk model in describing the global characteristics of small targets and analyze the model's limitation of strong bias to salient target signals.To alleviate the limitation,we propose a hierarchical maximal entropy random walk model.In order to enable the model to describe the local characteristics of small targets,we modify the weight matrix of the model and construct a confidence coefficient map based on the modified weight matrix to fuse the model's output distribution map,which further enhances the model's ability in expressing the local and global characteristics of small targets.Experimental results show that this algorithm has greatly improved the performance of target enhancement,background suppression,and multi-target detection accuracy compared with the existing algorithms.