Contour Detection Model Based On Target Depth Information And Receptive Field Dynamics

The human visual system has efficient architecture for information reception and integration for effectively performing visual tasks like detecting contours.In the human visual pathway,an external image first passes through ganglion cells and then through the lateral geniculate nucleus in its way to the visual cortex.Neurophysiological studies have found that lower neurons in the primary visual cortex converge in an orderly manner to form upper neurons,therefore,the simple cells in the primary visual cortex have the characteristics of the lateral geniculate nucleus.In order to further simulate the common processing mechanism of multiple visual information in the visual pathway,this paper introduces the binocular visual processing mechanism and the dynamic characteristics of receptive field into the contour detection task.The proposed model not only improve the performance of the contour detection model,but also in line with the visual information processing mechanism.It is a dynamic regulation process for the receptive field to receive external stimuli and the depth information is generated when the binocular cells receive information.At the same time,this paper considers that the visual system processes information in a parallel way in the shape,color and depth of the visual cortex,a contour detection model based on the characteristics of depth-sensitive cells and dynamic characteristics of receptive fields is proposed.The proposed model includes the processing of nature images and depth images.Nature image response was preprocessed using weighted least squares filtering;nature image and depth image responses were obtained by V1 neuron responses,and nature image and depth image responses were combined based on the optimal orientation.More importantly,depth information was used to calculate nature image and depth image responses to the receptive field size of each pixel.The experimental results based on the NYUD dataset,BSDS dataset,and MBDD dataset show that introducing depth information and dynamic characteristics of receptive fields into the contour detection model effectively improves the performance of the contour detection model.For the construction of the existing deep learning network model,most of them only consider the improvement of the VGG16 coding-decoding architecture.The deep learning model is inspired by biological visual,the paper proposes a contour detection model based on depth features and convolutional neural network.This model is mainly improved for the VGG16 preorder coding network.In this paper,the feature map obtained from the above biological mechanism model is integrated into the VGG16 coding network,and the feature map is also used as a side output layer and integrated with other side output layers.The experimental results based on the BSDS dataset and NYUD dataset show that the network model proposed in this paper effectively improves the accuracy of the contour detection model of deep learning,providing a new direction for the research of deep learning.