Object Detection Methods Based On Biologically Visual Mechanism

Object detection and recognition technology has attracted much research attention in recent years,lots of object detection and recognition methods have been proposed and successfully applied to many domains,like video surveillance,image compression and intelligent transportation,etc.Human visual information processing system can rapidly perceive stationary or motional targets from the complex environment.The human visual perception process seems simple,however,which involves complicated psychology and physiology mechanism.Until now,the machines' visual perception ability still falls far behind human visual perception system.In order to narrow this gap,with the help of the latest theoretical achievements in neurosciene and psychology,building the object detection and recognition method by imitating the perception process can largely improve the ability of machines' object detection and recognition.Inspired by the theory of the human visual attention mechanism,this paper implements a series of researches on object detection and recognition technology,and emphasizes researches on spatial salient object detection,spatiotemporal salient object detection,visual saliency computation applied to remote target detection and salient object detection based on visual multi-level perception mechanism.The details of the researches in following:At first,inspired by the theory of biological visual attention mechnism,we propose a laplacian Regularized kernel regression model for spatial salient object detection.The attention enhancement theory indicates that the diffusion of space-based attention generating object-based attention.Based on this theory,the result of classical eye fixation model is used as input,the laplacian Regularized kernel regression model is proposed for simulating the diffusion process of space attention to generate object-based attention.The experiment shows that the result after diffusion is much better than the initial eye fixation result.Notably,the final result achieves excellent performance on public datasets.In the second,inspired by two mutual channels structure of visual attention,a tensor-based spatiotemporal saliency computation model is proposed.The theory of two mutual channels shows: visual attention is firstly divided into ventral channel and dorsal channel,the two channels continually exchange information in the course of visual information processing,until finishing the target detection task.The division of the two channels is different,the ventral channel is mainly responsible for object appearance information,the dorsal channel is mainly responsible for the information of object location and motion.Based on this theory,we firstly use tensor to model image sequence,then,simulating the mutual process of the two channels through tensor computation to obtain final result of spatiotemporal saliency.The experiment results show that our model can achieve excellent performance in public datasets.In the third,in order to detect target quickly and accurately in complex and wide-range remote sensing image,inspired by visual attention selection mechanism,a fusion framework by combining line saliency with region saliency is proposed to detect salient object.The biological visual attention selection mechanism shows that: top-down visual modulating signal combining with bottom-up visual modulating signal generates final visual attention,the top-down modulating signal analyzing the given visual task by its prior knowledge to obtain the salient features of object,the bottom-up saliency is mainly obtained by visual feature contrast computation.Based on this theory,we firstly use bottom-up saliency model to obtain target candidates,then,convolution neural network uses its prior knowledge finishing the process of signal modulating to obtain final salient target extraction.The experiment results show that our model can achieve good performance.At last,in order to solve the problem that the scale and shape of object are diversified,inspired by the hierarchical visual perception structure,we propose a fusion method combining the low-level feature and high-level feature obtained by convolution neural networks to detect airplane and ship target.At the same time,we introduce deformable layer into convolution neural networks to solve the diverse scale and shape of target.Finally,based on the visual attention mechanism,the object contextual information is integrated into the network.The experiment results show that our model can achieve good detection performance.