Saliency Detection Of RGB-D Image Based On Multi-level Feature Supervised Fusion

Saliency target detection technology aims to identify the most attractive area in the scene by imitating human visual system to obtain valuable information.This technology has been widely used in many fields.In recent years,saliency detection algorithms for RGB-D images have been proposed continuously and achieved good results.Although the research on this technology has made great progress at present,there are still some deficiencies.This paper mainly proposes an improved depth learning model to solve the problem that the RGB-D image saliency detection algorithm can not effectively use the depth features,and the features can not be fully fused.The specific research contents are as follows:Aiming at the problem that the depth feature of RGB-D image saliency detection algorithm can not be effectively used and the feature fusion is insufficient,a multi branch backbone supervision network based on multi-level feature supervision fusion is proposed.Firstly,the network extracts multi-level features of RGB image and depth image based on Resnet50 network;Secondly,based on attention mechanism,depth improvement module is introduced to effectively mine depth features;Then,a feature group supervision fusion module is proposed.The mixed features after the fusion of the two features are input into the module in pairs from top to bottom for feature full fusion,and the upper group results and truth map supervision are added to the fusion;Finally,after iterative optimization,the last group of saliency maps is output as the result map.Experiments are carried out on widely used datasets,and compared with other detection models,the network achieves good results,with advantages and rationality.Aiming at the problems that the depth features of the RGB-D image saliency detection algorithm can not be fully utilized and the high-level and low-level features cannot be effectively fused,a depth learning network based on the multi-level feature supervised fusion is proposed from a new perspective.In this network,the Resnet50 network is used to extract the multi-level features of the two images;secondly,the side improvement module is introduced to optimize the feature denoising.At the same time,the depth information contribution calculation subnet is introduced to obtain the depth contribution map representing the net gain of the depth image to monitor the feature fusion;Then,based on the idea of inter-level supervision optimization strategy,the saliency map generated by the high-level feature fusion is used to guide the low-level feature fusion with its interval of one level;Finally,the saliency map of each group of outputs is given different weights according to the advantages and disadvantages,and the final result map is obtained through calculation.Through experiments on representative data sets and comparing with other advanced RGB-D image saliency detection models,it is shown that this network has superior performance and good robustness.