Salient Object Detection Via Multi-path Cascaded Deep Neural Networks

The attention mechanism of the human visual system allows humans to selectively focus on the most informative and characteristic parts of the visual stimulus,rather than the entire scene.This visual characteristic of human beings is learned by computers,which is called salient object detection.The traditional method can explore the saliency information in the natural scene images with single target or high contrast from many aspects by using the low-level visual information such as color and texture,etc.However,when it comes to some natural scene images with complex background,continuous target and low contrast,the traditional method usually has poor detection effect.Deep convolutional neural network can use a large number of neurons to learn multi-scale and multi-level feature information from the input image,which has more advantages than traditional method.Therefore,a variety of image salient object detection algorithms based on convolutional neural network are proposed in this paper.Its main contents are as follows::(1)A deep salient object detection model via attention mechanism and muti-path refinement is proposed in this paper.The model consists of two structures:multi-level feature extraction based on attention mechanism,and multi-path cascaded feature refinement.Multi-level feature extraction is based on squeeze and excitation residual network,and automatically learns and focuses on the input selectively.The multi-path cascaded feature refinement network has a flexible structure,and the saliency map can be obtained by multiple refinement,multiple up-sampling and multiple feature fusion of the extracted multi-level features.Compared with CHFR algorithm,MAE index decreased by1.75%,1.58%,1.03%and 1.23%respectively.The index of?-F_?improved by 5.92%,6.45%,10.23%and 5.09%,respectively.(2)A deep salient object detection model via context information and light-weight feature fusion is proposed in this paper.The model consists of three structures:multi-scale feature extraction based on context information,lightweight feature refinement and hierarchical feature fusion.Multi-scale feature extraction is based on Res2Net.The context information can be learned from the input image through the multi-scale receiving field at a fine-grained level,and the rough feature map can be generated.Then,the light-weight feature refinement model based on separable convolution is used to gradually refine the features from the high level to the low level.Finally,the multi-level features were fused to obtain a saliency map.Compared with Pool Net algorithm,MAE index decreased by 1.2%,0.4%,0%and 0.4%,respectively.Max-f measures increased by 0.4%,0%,-1.7%and 0.2%,respectively.(2)An image processing system is designed based on the model of salient object detection.The system has rich functions such as blurring of image background,matting with one key,style transformation and different space.In this paper,a variety of salient object detection algorithms based on convolutional neural network are proposed,which not only improve the detection performance,but also reduce the detection cost.The prototype of image processing system based on salient object detection is designed,which has a friendly human-computer interactive interface,and can bring a lot of convenience to users who have requirements for retouching.