PriorNet: Two Deep Prior Cues for Salient Object Detection
Ge Zhu, Jinbao Li, Yahong Guo
Abstract
Current salient object detection methods achieve good performance by aggregating multi-level features from fully convolutional network. However, in the process of feature aggregation, the noise will be introduced due to the information difference between different level features. Besides, the semantics of high-level features will be diluted as they pass on the top-down pathway, which makes it difficult for model to separate the salient objects from background completely in complex scenes. To address the above problems, we propose two deep priors, including global location prior (GLP) and local contrast prior (LCP). The GLP is generated from prediction, which can enhance the semantics of aggregated features at each level to locate salient objects. Compared with aggregation-based models that directly use high-level features as enhanced semantics, the proposed GLP contains richer semantics and details. The LCP is inferred based on the weighted differences between center pixel and surrounding pixels in backbone features, which can select discriminative features and suppress the noise from aggregated features by multiplication with residual connection. Based on the two priors, we propose a novel twice-decoding network, where the first decoding is to generate GLP by aggregating multi-level features and LCP, and the second decoding is to refine salient objects by using GLP and LCP. Different from previous methods which use a recurrent structure to merge output into input images, the proposed network only applies the output in decoding to avoid interference of raw images. Comprehensive experiments on five datasets show that the proposed method outperforms state-of-the-art ones on five evaluation metrics.