Research And Application Of Target Perception Calculation Model

The human visual system has a strong ability to perceive image information.The complex visual mechanism is the prerequisite to support the efficient operation of the human visual system.Therefore,studying the visual perception mechanism is of great significance for understanding the target perception process.In recent years,the rapid development of deep learning technology that simulates the working mode of human brain neurons has contributed more important new ideas to the research of target perception computing models.Inspired by this,this article first considers the processing process of visual information flow by the retina,lateral geniculate body,nerve synapse and visual cortex in the main visual pathway of the visual system,and constructs a target contour detection model based on the visual perception mechanism.;Then further consider the problem of regional target perception,build a visual saliency detection model based on the convolutional neural network,and combine the established discriminator and filter to achieve the purpose of expanding the training data set,effectively improving the saliency of the original network calculation model The effect of sex detection;finally,the saliency detection model is transferred to the task of pedestrian reidentification.The main research of this article is as follows:(1)A contour detection method based on the hierarchical response model of the main visual path structure is proposed.First,according to the dark field characteristics of the retinal photoreceptor cells and the orientation selectivity of the multi-scale classical receptive field,the detection path of the advanced contour and the global contour is constructed;the texture sparse coding of the information is simulated by the characteristics of the LGN cell,combined with the side inhibition effect of the non-classical receptive field Inhibition of strong background texture;In addition,a micro-motion integration mechanism is proposed in the LGN area to reduce texture redundant information,and then achieve information related transmission through adaptive synapses;finally the primary contour response is forwarded to the V1 area across the visual area and corrected by the global contour Then,it is quickly fused with the advanced contour response to realize the contour detection of the target subject.Experimental results show that this method can distinguish contours from texture edges more effectively,highlight the contours of the target subject and suppress irrelevant background noise.(2)A visual saliency detection enhancement method based on adversarial training and data screening is proposed.First use the ENet network as the saliency detection framework;build a discriminator to discriminate pairwise between the training data set and the new data set through the output features of the ENet decoding module,during which ENet and the discriminator alternate against training;then use the CAM algorithm to find images of interest After the region,the un ROIErase method is proposed to train the classification module,which forms a data filter together with the feature distance measurement module;finally,the reliable image and pseudo-labels are filtered from the new data set and expanded to the training data set,and the ENet network is retrained for visual saliency detection.The experimental results show that the above method can expand effective new data for the source data set,drive the network to mine more diverse feature distribution information,improve the saliency detection ability and generalization performance of the network model,and effectively alleviate the network overfitting problem.(3)A method of applying saliency detection to pedestrian re-identification is proposed.First,use the trained ENet network to obtain the saliency map of pedestrians,and then build a refining layer at the rear of layer3 of Res Net50 to extract features,and expand the feature map to the same size as the saliency map through bilinear interpolation,and then merge with the saliency map to enhance The response of local salient features;Finally,after the global features and local features are spliced,the model is trained by means of joint loss and label smoothing to improve the performance of the model.The experimental results show that applying the pedestrian saliency map to pedestrian re-recognition can improve the quality of local salient features,and the method of co-training it with global features can improve the re-recognition ability of the network.