Universal Target Detection Based On Multi-dimensional Local Adaptive Regression Kernel

In recent years,computer vision has received more and more scientists' attention.Among them,target detection is one of the most important parts of computer vision,which is divided into two categories:Supervision and Non-supervision.Traditional supervised methods are all based on classifier.Because training classifier needs a lot of data and constantly adjusts training parameters,it causes large time cost.To solve these problems,the unsupervised methods are a good way to follow.This paper,analyses the local adaptive regression kernel(LARK),and proposes two robust target detection models,which are applicable to similar targets of different shapes or scales(i.e.generality).(1)A target detection model based on Gaussian differential LARK.This paper studies the feature extraction operator of LARK deeply,and considers the suppression characteristics of the non-classical receptive field of human visual neurons,and constructs the feature of Gaussian differential LARK(GLARK).At the same time,3D-GLARK feature operator is built by combining with the context information of image sequence to constraint target location,solve weak-edge targets in complex scene detection problem,improve the accuracy of target detection.(2)A target detection model based on multidimensional LARK.Based on the theory of the above model,further combined with the abundant information on the image spectrum,this model ensures the integrity of the target structure;at the same time,deep learning algorithm is introduced into this model to enhance the adaptive ability of the target indication window;in order to solve the problem of low efficiency of the LARK target detection method,this model designs parallel distance scoring mechanism through parallel processing similarity calculation process.Experiments show that the GLARK operator has strong perception ability for weak edges of targets,and has good detection results for some complex scene videos of such qualities,partial occlusion and background interference.broadening the LARK recognition width to the spectral domain,combined with deep learning greatly,multidimensional LARK target detection model is designed to reduces the error detection rate and false negative rate of traditional LARK,and resolves the discomfort problem of marking the target window size;parallel distance scoring mechanism which greatly improves the detection efficiency of this method has made important paving for real-time detection.