Research On Deep Learning For Light Field Saliency Detection

As a basic and challenging dichotomous task,salient object detection task has attracted the attention of many scholars.Its purpose is to detect the most interesting targets in a scene or area.Traditional salient object detection based on manual feature extraction has great limitations because it cannot obtain global context semantic information.However,the rise of convolutional neural network has greatly improved the performance of salient object detection,which is mainly because of its automatic extraction and combination of high-level and local low-level features.Based on different input data,salient object detection is mainly divided into three categories: 2D,3D and 4D Light field saliency detection.This paper mainly focuses on 4D light field salient object detection and conducts the following research work.For 4D light field salient object detection,there are only 100 publicly available LFSD images in the light field saliency data set,which severely limits the development of the 4D method.It makes it difficult for the data-driven deep learning method to effectively learn representative features.Therefore,in view of insufficient light field data,this paper introduces the largest light field salient object detection data set(DUT-LFSD)up to now,including 1462 all-focus RGB images,depth maps,focal stack images and corresponding the ground truths maps.And this data set makes the image closer to the life scene because it is shot in the real world.Meanwhile,it also includes many challenging scenes,such as similar foreground and background,complex background,transparent objects and multiple targets.This makes the proposed data set sufficient to verify the validity and robustness of the model.In addition,this paper proposes a novel of Light Field fusion network for saliency detection,using two-stream networks of feature extraction to extract raw RGB and light filed features,and then use Light Field Refinement Module(LFRM)inspired by residual learning thoughts to the extraction of the Light Field information.Deep supervision mechanism is used to improve the learning effectiveness of network and accelerate the convergence of the network.Then,a light-field Integration module(LFIM)is designed to weight the contributions of each light field integration feature,and learn the semantic relations within the integrated feature to accurately locate and identify salient objects.A large number of experimental comparisons are made in the proposed data set and the publicly available data set,and the experimental results consistently prove that the proposed method is superior to all the existing 2D,3D and 4D light field salient object detection algorithms.