Improved cascaded partial decoder for boundary aware salient object detection

Abstract

Compared with low level features, high level features, have greater spatial resolution and less performance contribution, but are more computatively expensive. At the same time, most of the previous researches focused on the regional accuracy, and the boundary mass was less studied. In this paper, we improve the CPD framework to use a new hybrid loss for boundary aware salient object detection. Method of this paper integrating the features of the deeper layer to obtain relatively accurate salient maps. And using hybrid loss, the improved CPD framework can effectively conduct significance detection to obtain clear boundaries. Experiments on six benchmark data sets show that the proposed method not only runs faster than the existing model, but also performs well in accuracy.