A General Deep Saliency Enhancement Framework
Lei Bao, Yunfei Zheng, Xiaoyan Qin, Haiqiang Dong · 2020
Identifying more accurate boundaries of salient objects has attracted broad attention in the field of saliency detection. To solve this problem, most of the previous works usually choose to train an extra edge detection branch or take a complicated loss to guide the training process. Different from these methods, this paper presents a general deep saliency enhancement framework which takes an unsupervised method to append the salient edge information. Firstly, trained deep-learning based saliency detection method works as top-down process, which is in charge of detecting main parts of salient objects. The obtained saliency region is further used as a rough mask to guide the following process. Then an unsupervised saliency edge detection method based on artificial characteristics works as bottom-up process, which is in charge of saliency map refinement. The bottom-up mechanism is used for saliency boundary estimation, while human visual mechanism is simulated to detect distinguishable saliency. After that, the rough saliency masks and saliency boundary detection results are fused by deep fusion method to output the final saliency maps. Experiments show that, not only this framework could achieve state-of-the-art performance on target detection, but also give more useful detailed inner information of salient objects.