Recently, student SONG Dawei from the Xi'an Institute of Optics and Precision Mechanics (XIOPM) of the Chinese Academy of Sciences (CAS) proposed a hierarchical edge refinement network (HERNet) for accurate saliency object detection. Their up-to-date result was published on IEEE TRANSACTIONS ON IMAGE PROCESSING.

Given an image, the goal of salient object detection is to locate interesting targets that look quite different from their surroundings. Compared to most existing convolutional-neural-network methods which usually ignore the blurred edges of salient objects, the HERNet achieves accurate salient targets detection.

Accurate saliency object detection is an important task of computer vision. To meet this demand, Song and his team members proposed a novel network. The whole structure was disassembled into two significant modules, saliency prediction network and edge preserving network. And the model was trained by three necessary supervisions, structure supervision, hybrid supervision, and edge supervision.

The architecture of HERNet. (Image by XIOPM)

In summary, the paper empirically showed edge blurring of prediction map was a challenging task in salient target detection, and the proposed HERNet can effectively mitigate it. Comprehensive experimental results demonstrate the superiority of HERNet under different evaluation metrics. In the future, the proposed method will inspire the designing effective architecture of accurate salient target detection.