Salient object contour extraction based on pixel scales and hierarchical convolutional network

Abstract

Image salient object contours are helpful for many advanced computer vision tasks, such as object segmentation, action recognition, and scene understanding. We propose a new contour extraction method for salient objects based on pixel scale knowledge and hierarchical network structure, which improves the accuracy of object contours. First, a deep hierarchical network is designed to capture rich feature details. Then, a new loss function with adaptive weighted coefficients is developed, which can reduce the uneven distribution influence of contour pixels and non-contour pixels in training datasets. Next, the object contours are classified based on the scale information of contour pixels. By importing the prior knowledge of scale categories into the network structure, the model requires a small number of training samples. Finally, the regression task of contour scale prediction is added to the network, and the precise contour scales of foreground objects are performed as prior knowledge or an auxiliary task. The experimental results demonstrate that compared with related methods, the proposed method achieves satisfactory results from precision/recall curves and F-measure score estimation on three datasets.