Infrared Dim Target Detection In Complex Scene

Infrared dim target detection is one of the core technologies of modern advanced weapon systems.It is also widely used in the fields of industry,transportation,security,medical and other civil fields,and becomes a hot research topic of computer vision.The characteristics that small imaging area and few features make dim target detection vulnerable to background clutter.Therefore,how to detect small targets in complex scenes accurately,reliably and quickly is becoming a technical difficulty and research hotspot.Based on the practical needs,dim infrared targets detection algorithm in complex scenes is studied in this paper,and experiments based on the real scene image sequences are implemented to verify the feasibility and effectiveness of these algorithms.In this paper,the three key components of infrared images,including target,background and noise,are analyzed and modeled firstly.On this basis,the scale invariant dim target detection algotithm based on DoG scale space and signal-to-clutter ratio optimizing is prezent.In order to solve the problem of high computation complexity and realize fast detection,some improvement measures are proposed: replacing DoG operator with CenSurE operator to quickly calculate the multiscale filtering response,and rasterizing the image and then searching the maximum response in grid to reduce the calculation.It is difficult to detect dim and weak target because of its weak energy and low contrast ratio.So the method of spatio-temporal energy accumulation is used to enhance the contrast between target and background,and then improve the signal-to-noise ratio of the image and target detection probability.To make full use of target's frame correlation,two algorithms,a moving weighted pipeline filtering algorithm based on target motion estimation and a moving target detection algorithm based on optical flow,are introduced.Both two methods are effective in suppressng false alarms,and the experimental results show that the algorithm based on optical flow works better than the updated pipeline filtering algorithm.Traditional IR small target detection mainly depends on filtering algorithms.In this paper,the dim target detection is considered as a non-equilibrium classification problem.To study the feasibility of using machine-learning technology to solve this problem,a statistical feature based on histogram distribution is proposed to describe the target.Moreover,the classifier is trained based on cost-sensitive Gentle Adaboost algorithm to classify target and non-target.The experiments are carried out using the real scene infrared image sequences,firstly the multi-scale CenSurE operators are used to locate the local maximum,then the image subblocks are automatically extracted as the candidate targets,finally the classifier trained by the Gentle AdaBoost algorithm determines the target and non-target.The K cross validation results show that the algorithm is feasible and effective.