Fully Convolutional Networks For Image Saliency Detection

The human visual system can quickly filter out the most valuable information from complex scenes and then focus on the areas of interest.Image significance detection is to extract important areas of the image through computer simulation of human vision system,so as to simplify the complexity of subsequent utilization.According to existing studies,significance detection methods can be heuristic based or task-based based on learning.In this paper,the significance detection algorithm based on learning is studied,and different full convolution neural network architectures are proposed to realize more robust significance detection model to adapt to the complex background environment.This paper mainly studies the following aspects:(1)A new significance detection method based on fusing all convolutional layer features is proposed.First,all convolution features are mapped to internal multiple scales.Then,adaptive combination of each scale feature is used to predict the significance graph of each scale.At the same time,the significance graph of each scale is fused to produce the fused significance graph.Finally,the final significant graph is obtained by further optimization of the random field with full connection condition.The algorithm is verified in the open data set,and the results show that the method can effectively fuse all levels of features,and the significant region is more prominent and the significant boundary is more accurate.(2)Aiming at the problems of the current significance detection algorithm based on deep learning,which detects significant objects with mixed background noise and fuzzy edge areas,a new significance detection method based on cascade full convolution neural network is proposed.The network is mainly composed of a two-level full-convolutional neural network.In the first stage,a coding-decoding full-convolutional neural network is constructed.A pyramid module is introduced into the coding network to obtain global information of different regions and suppress the interference of background noise.In the second stage,the edge detection network is designed to extract the local semantic information of the network and learn the edge information of the significant region.By fusing the two-stage significance graph,the accurate boundary significance graph is obtained.The proposed algorithm is verified on two open data sets,and the results show that the proposed algorithm has better performance for complex boundary information and can effectively suppress background noise.(3)Aiming at the current significance detection algorithm based on deep learning that is difficult to detect the robust significance region in complex scenes,a new significance detection method based on full convolution neural network guided by prior features is proposed.This method can be combined with significant prior knowledge to make more accurate prediction.In addition,prior features and deep features are effectively fused through the feature fusion method of attention mechanism,and finally the previous errors are corrected iteratively through the cyclic decoding network,and the significance region is automatically learned to improve,so as to produce more reliable final significance graph prediction.Experimental results on two open data test sets show that the proposed algorithmis robust in complex scenarios.