Bi-Connect Net for salient object detection

Abstract

As a challenging task for pixel-wise image analysis, salient object detection has made huge progress in recent years. However, there still exists a difficult problem: detection of distinguishing a salient and non-salient object in multiple objects under complex background (e.g. blur, translucent, light reflection, etc.). Our proposed method cast this difficulty as information dissolve problem in deep convolutional network, which is manifested as: first, the model cannot grab whole details of a salient object at training phrase; second, due to the isolation between layers and blocks, the valued information is blocked within a block, which leads to the difficulty in obtaining the position and the edge of salient objects simultaneously; third, the output of the network is a low-resolution saliency map, which cannot accurately express the edge of salient objects. To address information dissolve problems, we construct a Bi-Connect Net (BCN) composed of forward connection subnet and reverse side connection subnet. Besides, the proposed adaptive learning fusion method not only stress all blocks contribution but also combine multiple features with different scale, so that grab more details on the right salient location and precise edges at the same time. Extensive experiments show that our proposed Bi-Connect Net can outperform the state-of-the-art methods.