Research On Visual Saliency Detection Algorithm Based On Multi-scale Deep Feature Fusion

In the era of big data,the image data that people are exposed to has increased dramatically,and how to extract effective information from images has important practical value.In order to enable the computer to have the ability of human vision to quickly find conspicuous objects in vision,many researchers have been committed to the research of visual saliency detection algorithms.The visual saliency detection algorithm aims to simulate the human visual attention mechanism and detect the most attractive salient target in the image.It is currently widely used in image segmentation,object detection,object tracking and other fields.With the development of deep learning,visual saliency detection has made a qualitative leap.How to efficiently fuse convolutional features,solve the problems of edge blur,and build efficient models has gradually become a research hotspot.This thesis mainly studies the visual saliency detection algorithms of images based on multi-scale deep feature fusion.The research contents are as follows:(1)To solve the problem that current saliency detection algorithms are difficult to effectively integrate multi-scale features and edge blur,a saliency detection method that gradually integrates multi-scale contextual features from top to bottom is proposed.First,a multi-receptive field module(MRFB)is designed to further capture multi-scale contextual information.Then,a feature fusion module is used to progressively fuse different outputs of MRFBs from top to buttom,which can effectively filter out the non-complementary parts of the high-level and low-level features.Afterwards,the refined residual block is applied to further refine the results.Finally,the edge-aware loss is used as an aid to guide the network to learn more sharpened details of the salient objects.The whole network is trained end-to-end without any pre-processing and post-processing.Evaluations on six benchmark datasets demonstrate that the proposed algorithm has achieved a relatively advanced level and the boundaries of salient objects are more accurate.(2)To solve the problem that the accuracy and real-time performance of current saliency detection algorithms are hard to balance,a novel globally guided multi-scale feature aggregation lightweight network for saliency detection is proposed.First,the lightweight classification network Mobile Net V3 is used as the backbone network to extract different levels of features of the image.Then,the lightweight multi-scale receptive field module is applied to further enhance the global representation of the highest level feature extracted by Mobile Net V3.Finally,under the guidance of high-level semantic information,the progressive feature fusion module is utilized to progressively fuse multi-level and multi-scale features from top to bottom,and the commonly used cross-entropy loss function is used for multi-stage deep supervised learning to obtain the final optimization result from coarse to fine.The whole network is an absolute end-to-end framework without any pre-processing and post-processing.Exhaustive experiments conducted on six benchmark datasets show that the proposed method is significantly better than the comparison methods.At the same time,the model is only about10 MB and run at a speed of 59 FPS on a GTX2080 Ti GPU when processing a 400×300 image.