Research On Image Instance Segmentation Based On Fractional Nonlinear System And Reinforcement Learning

At present,image instance segmentation is a research hotspot in computer vision.Its purpose is to identify specific types of instances in a given image,and finally segment the masks of each instance.However,the image processing capabilities of computers are still far away.Does not reach the level of biological vision.Therefore,from the perspective of simulating the biological visual nervous system,based on the theory of nonlinear chaotic synchronization control and reinforcement learning,this paper proposes a model that combines neurophysiological characteristics and image instance segmentation,and has achieved certain results on a specific type of standard data set.The effect of this kind of brain function bionic theory has broad application prospects in artificial intelligence.The main work of this paper includes the following two aspects:First,in view of the problem that the existing image instance segmentation model does not fully simulate the information processing method of the human brain visual perception system,an image instance segmentation AISM model combining nonlinear chaotic phase synchronization and reinforcement learning is constructed.Through the study of the "where" flow and the "what" flow in the neural processing mechanism of the brain,the initial construction of the "what" flow is carried out with the non-linear chaotic phase synchronization image feature binding and segmentation method,and then the continuous decision-making method in the reinforcement learning is used Learn and reason about the segmentation results of the nonlinear chaotic phase synchronization network,complete the optimization of the "what" stream and the positioning of the "where" stream,and finally segment the accurate image instance contour curve.Experimental results in specific categories of Pascal VOC 2007 and Pascal VOC 2012 data sets show that this method can effectively perform instance segmentation and far exceeds the nonlinear benchmark model.Second,in view of the problem that nonlinear network does not fully use the information of various layers of the image and the reinforcement learning is difficult to process high-dimensional images,the information receiving mode and processing mechanism of the visual nervous system are analyzed from another perspective,and a joint image chaotic attractor is constructed.And reinforcement learning image instance segmentation FCRL model.This research first puts forward the image layering theory and the concept of image chaotic attractor,and then simulates the image information recognition process of the pixel-region level "what" flow and "where" flow in a non-linear way,and finally the region-total "where" The information processing process of "flow" and "what" flow performs Markov modeling to complete the segmentation of image target instances.The experimental results on the Pascal VOC2007 and Pascal VOC 2012 standard data sets of specific categories show that compared with the previous fractional method,this model effectively improves the detection accuracy;compared with the AISM model,it greatly accelerates the detection speed.