Expanded Edge Penalty Loss for Salient Object Detection

As an indispensable preprocessing technique for image understanding, salient object detection aims to extract interesting regions from an image for subsequent processing, which has attracted much attention since its wide range of applications. Recently, with the rapid development of artificial intelligence and machine learning, deep neural network especially deep convolutional neural network based methods have achieved competitive performances because of their strong feature representation capability. However, most of these methods often suffer from coarse boundaries. The main reason is that equal penalty factor is applied to each pixel in the image to optimize the network, but there are huge distinctions in prediction complexity among different ones actually. Specifically, pixels closer to the boundaries are increasingly difficult because of their huge gaps between structural information and semantic label. In these cases, we present an Expanded Edge Penalty Loss (E
²
PL) for salient object detection. E
2
PL gives bigger penalty factors to pixels distributed in boundary and near-boundary regions, and further dynamically adjusts their contributions to the model optimization. In addition, the experimental results on five public and challenging datasets have validated the superiority and effectiveness of the proposed method.