Research On Image Saliency Detection Based On Deep Model

The purpose of image saliency detection is to find the most attractive target in natural images,and provide convenience for the following image processing.It has been widely used in many tasks,such as visual tracking,image retrieval and semantic segmentation,etc.,Existing image saliency detection methods have achieved good detection results in most natural scenes,but there are still some problems,like low accuracy,incomplete target areas and blurred boundaries.In response to these problems,this paper proposes two saliency detection algorithms and implements a saliency detection system on this basis.The main research tasks of this paper are as follows:To solve the problem that most saliency detection methods only consider local contrast and ignore global information which result in low accuracy,a saliency detection algorithm based on the fusion of local and global features is proposed.Firstly,the saliency value is calculated by combining the center prior,target prior,dark channel prior information and local contrast of the image,then a local saliency map is generated through the strategy of multi-scale fusion.Secondly,A fully convolutional neural network model is constructed to achieve the prediction of image pixel saliency values,and also generate a global saliency map.Finally,the above two saliency maps are combined as the final saliency map.Experimental results show that the proposed algorithm can get higher F-measure and lower MAE value,and at the same time it can obtain more accurate saliency area.To solve the problem that the saliency targets detected in complex scenes have incomplete target areas and blurred boundaries,a saliency detection algorithm based on multi-level feature fusion is proposed.Firstly,the multi-level feature information is extracted,and the receptive field module is embedded to obtain rich context feature information.Then,the multi-level feature fusion is performed to ensure the integrity of the saliency detection area.Finally,a boundary optimization module is constructed to optimize the boundary of salient targets and output the final saliency map.The experimental results show that the salient regions obtained by the proposed algorithm are more complete and the edge of the target is clearer.In the comparison of various evaluation indicators,the proposed algorithm has achieved good performance.On the basis of the above work,an image saliency detection system is implemented based on PyQt5 graphical programming technology.Given an input image,the system can perform saliency detection operation to generate corresponding saliency maps,thereby enhancing the operability of saliency detection.