Research On RGB-D Image Co-Salient Object Detection

The visual attention mechanism allows people to automatically capture the most attractive part of the complex scene.As a branch of the field of computer vision,salient object detection is devoted to simulating the attention mechanism so that the computer can automatically discover salient objects in the scene.It has been a vigorous research topic in the past decade and has been applied to segmentation,redirect,enhancements,foreground annotations,and many other visual tasks.With the explosive growth of data volume in recent years,multiple related images need to be processed simultaneously by people.Co-salient object detection is an extension of traditional single image salient object detection.Its purpose is to find the common and salient objects in multiple related images.Unlike the single image salient object,co-salient object need to be salient not only in their respective images,but also in other related images.Previous research on salient object detection mainly focused on RGB images,only using color information and ignoring depth information,but depth information has proved to be useful in many computer vision tasks.This article will use the depth information to supplement the RGB color information,and perform co-salient object detection on a group of related RGB-D images.The main work is as follows:(1)In reality,many image groups containing the same or similar object are collected in similar scenes,which will lead to the problem of incorrectly detecting the same background area when detecting the same salient area and in order to better represent the image Information,a RGB-D co-salient object detection method based on the composition and feature optimization is proposed.First,the method of existing RGB-D image salient object detection is used to generate an initial saliency map.Then the initial saliency map is used to obtain the initial foreground area set and the image edge area set of the image group,and based on the two sets,a closed loop graph structure is constructed with superpixel as node for the original image group,and select the feature set composed of the color,texture and different levels of high-dimensional semantic features extracted from the original image through genetic algorithm to obtain a feature subset suitable for the image group.Then the initial co-saliency map is obtained by combining the closed-loop graph and the feature subset calculation.Finally,the depth map is used as an aid to optimize the initial co-saliency map through an energy function to generate the final co-saliency map.(2)Some images in a group of images are simple and some images are complex,and simple images are easier to detect the foreground area,so this paper uses simple images that are easy to detect to guide complex images to perform co-salient object detection,and proposes a RGB based on sample selection RGB-D co-salient object detection method.First,based on 4 scoring indicators,selecting simple images with some salient object that are easy to detect from the image group to form a simple image set;then,based on the principle of collaborative consistency,extract positive and negative samples from the simple image set and using the deep learning model to extract the high-dimensional semantic features represent positive and negative samples;furthermore,the co-saliency classifier trained with positive and negative samples is used to classify the superpixels in the image to obtain the co-saliency object area;finally,after a smooth fusion operation,the final co-saliency map is obtained.The two methods proposed in this paper are compared with several other methods on two public data sets,and the experimental results show that the methods proposed in this paper have better performance.