Cognitive And Neural Modeling For Visual Information Representation And Memorization

Human can easily recognize objects no matter how complex the background is,and this ability is mainly supported by the powerful visual perception and memory system.How to extract useful information from a large number of visual information,and how to effectively organize and represent these information has been a hot and challenging topic in the field of computer vision.Although a great number of achievements have been made in the feild of bio-inspired perception system,however,some of the proposed perception models are too complex,and some of the biological characteristics are not expressed enough to play important role in biological visual information processing.So far,little is known about the neurophysiological mechanism of human brain memory,and less research is focused on learning and memory of natural image.Therefore,it is of great academic and practical value to study how visual information is represented,stored and memorized in the brain and to model the cognitive neural mechanism.The main work of the paper is as follows:1.On the basis of the research of biological visual perception mechanism and visual cognitive modeling,in order to better simulate the biological characteristics of nerve cells in V2 and V4 areas of human visual pathway,the multi-firing K-means and non-negative sparse coding(NNSC)were introduced into the HMAX model to develop a five level biological visual perception model,denoted as Sparse-HMAX.Multi-firing K-means was used to simulate the property of V2 neurons.By introducing NNSC into S2 layer,the proposed model could better explain the response characteristics of V4 neurons.Experimental results show that the proposed model outperforms the HMAX model in classification accuracy and time efficiency,and is superior to similar models in most cases.2.In view of the fact that HMAX model is sensitive to rotation and does not take into account the characteristics of neurons in the retina and lateral geniculate,and has the limitation of inadequate description of biological characteristics of higher neurons,a bio-inpired model combined with Log Polar Gabor(GLoP)filter and multi-manifold sparse coding was developed by modeling the neural mechansim of visual cortex.Contrast normalization was used to pre-process the visual input,and GLoP filter was used to simulate the simple cells in V1 area.Considering the sparse firing nature of V4 neurons and the way of manifold visual perception,multi-manifold sparse coding(SCMM)was used to emulate cells of the V4 area.In addition,a template learning method based on multi manifold dictionary learning was proposed.Experimental results show that the proposed model shows a significant advantage over the HMAX model,and the robustness to rotational transformation is improved greatly.3.To realize rapid storage and retrieve of visual information,an Increment Pattern Association Memory Model(IPAMM)was proposed based on the research findings of cognitive neuroscience.Leabra learning mechanism was introduced into pattern association network to replace the simple Hebb learning rules,and incremental learning can be achieved by assigning and different separate weights to different patterns.Experimental results show that the proposed model can perform image classification task well,and has good recognition performance in the case of small training samples.Meanwhile,the proposed model has high time efficiency and can better meet the requirement of practical applications.4.In view of the fact that Friston's theory of free energy provides a theoretical framework for the perception and memory of human brain,and there is no practical application at present,we proposed a visual memory model based on free energy theory and Restricted Boltzmann Machines(RBM).The system's free energy was minimized based on RBM model,so that the visual information learning and extraction can be realized.Experimental results show that the proposed model is superior to SVM and SOINN models in image classification,and has equivalent classification performance as IPAMM model.5.In order to realize rapid detection and identifition of crop pests,the proposed visual perception model combined with visual attention mechanism was applied to detection and recognition of crop pests.First,pests were detected and regions of interest are extracted based on the Saliency Using Natural Statistics(SUN)model.Then invariant features were extracted based on the proposed visual perception model and the LCP(Local Configuration Pattern)algorithm.Finally,SVM was used to classify and recognize the pest.Experimental results show that the proposed method can accurately detect the target area in complex natural environment,and has good recognition performance.It provides a new idea and method for rapid detection and identification of crop pests.