Salient Object Detection in Low Contrast Images via Global Convolution and Boundary Refinement

Nan Mu, Xin Xu, Xiaolong Zhang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 0-0

Benefit from the powerful features created by using deep learning technology, salient object detection has recently witnessed remarkable progresses. However, it is difficult for a deep network to achieve satisfactory results in low contrast images, due to the low signal to noise ratio property, thus previous deep learning based saliency methods may output maps with ambiguous salient objects and blurred boundaries. To address this issue, we propose a deep fully convolutional framework with a global convolutional module (GCM) and a boundary refinement module (BRM) for saliency detection. Our model drives the network to learn the local and global information to discriminate pixels belonging to salient objects or not, thus can produce more uniform saliency map. To refine the localization and classification performance of the network, five GCMs are integrated to preserve more spatial knowledge of feature maps and enable the densely connections with classifiers. Besides, to propagate saliency information with rich boundary content, a BRM is embed behind each convolutional layer. Experiments on six challenging datasets show that the proposed saliency model achieves state-of-the-art performance compared to nine existing approaches in terms of nine evaluation metrics.