Research On Image Saliency Detection Algorithms Based On Visual Perception

Image saliency detection is an important research problem in computer vision.It can detect and segment the most attractive objects in an image and is used as a preprocessing technology for many computer vision tasks.With the development of deep learning technology,many algorithms based on convolutional neural networks emerge and their performance has been greatly improved compared with traditional methods.However,it remains a challenge to effectively utilize multi-level features in convolutional neural networks.Improper feature extraction and fusion may lead to incomplete salient objects or background interference.In addition,some studies introduce the edge information of salient objects into saliency detection,which may lead to internal missing of salient objects or interference of background boundary due to inaccurate extraction or underutilization of edge features.Therefore,how to use edge information to assist saliency detection is another challenge.Based on the above challenges,this paper studies the algorithms based on convolutional neural network from the perspective of simulating human visual perception ability to improve the localization and boundary accuracy of salient objects,as follows:(1)This paper proposes a saliency detection algorithm based on a progressive dualattention residual network,which uses two complementary attention maps to guide the residual learning,thus refining the prediction in a coarse-fine way.In this algorithm,a dual attention residual module is designed to make use of the foreground attention map and its corresponding background attention map,so that the network can learn the residual details from the salient and non-salient regions.In addition,a hierarchical feature screening module is designed to obtain more powerful global contexts by enhancing the information interaction and feature representation among multi-scale features.Experimental results show that the proposed algorithm is superior to 18 advanced algorithms on 5 datasets,which proves the effectiveness of the proposed algorithm.(2)This paper proposes an edge-aware saliency detection algorithm based on global and local information aggregation.The algorithm firstly proposes a global guidance module,which is composed of a global feature discrimination module and a local feature discrimination module.The former exploits the inter-channel relationship of global semantic features to boost representation power,and the latter enables different side-outputs to generate discriminative local features by fusing with global attentive features.Then,an edge-aware aggregation module is proposed,which uses the correlation between salient edge information and salient object information to generate predictions with definite boundary.The experimental results show that the proposed algorithm is compared with 17 advanced algorithms,and achieves good results on 6 datasets.