Lightweight adversarial network for salient object detection

Abstract

Recent advance on salient object detection benefits mostly from the revival of Convolutional Neural Networks (CNNs). However, with these CNN based models, the predicted saliency map is usually incomplete, that is, spatially inconsistent with the corresponding ground truth, because of the inherent complexity of the object and the inaccuracy of object boundary detection resulted from regular convolution and pooling operations. Besides, the breakthrough on saliency detection accuracy of current state-of-the-art deep models comes at the expense of high computational cost, which contradicts its role as a pretreatment procedure for other computer vision tasks. To alleviate these issues, we propose a lightweight adversarial network for salient object detection, which simultaneously improves the accuracy and efficiency by enforcing higher-order spatial consistency via adversarial training and lowering the computational cost through lightweight bottleneck blocks, respectively. Moreover, multi-scale contrast module is utilized to sufficiently capture contrast prior for visual saliency reasoning. Comprehensive experiments demonstrate that our method is superior to the state-of-the-art works on salient object detection in both accuracy and efficiency.