Saliency Contour Extraction Based on Multi-level Feature Channel Optimization Coding

Abstract

A biomimetic vision computing model based on multi-level feature channel optimization coding is proposed and applied to image contour detection, combining the end-to-end detection method of full convolutional neural network and the traditional contour detection method based on biological vision mechanism. Considering the effectiveness of the Gabor filter in perceiving the scale and direction of the image target, the Gabor filter is introduced to simulate the multi-level feature response on the visual path. The optimal scale and direction of the Gabor filter are obtained based on the similarity index, and they are used as the frequency separation parameter of the NSCT transform. The contour sub-image obtained by the NSCT transform is combined with the original image for feature enhancement and fusion to realize the primary contour response. The low-dimensional and low-redundancy primary contour response is used as the input sample of the network model to relieve network pressure and reduce computational complexity. A fully improved convolutional neural network model is constructed for multi-scale training, through feature encoder to feature decoder, to achieve end-to-end pixel prediction, and obtain a complete and continuous detection image of the subject contour. Using the BSDS500 atlas as the experimental sample, the average accuracy index is 0.85, which runs on the device CPU at a detection rate of 20+ FPS to achieve a good balance between training efficiency and detection effect.