采用跳层卷积神经网络的RGB-D图像显著性检测  

RGB-D Image Saliency Detection Using Skip-Layer Convolutional Neural Network

作　　者：陈曦涛 訾玲玲[1] 张雪曼 CHEN Xitao;ZI Lingling;ZHANG Xueman(School of Electronic and Information Engineering,Liaoning Technical University,Huludao,Liaoning 125105,China)

机构地区：[1]辽宁工程技术大学电子与信息工程学院,辽宁葫芦岛125105

出　　处：《计算机工程与应用》2022年第2期252-258,共7页Computer Engineering and Applications

基　　金：国家自然科学基金(61702241,61602227)。

摘　　要：RGB-D图像显著性检测旨在提取三维图像中的显著目标。为解决当前显著性检测算法难以检测出光线干扰场景内的目标和低对比度的目标等问题,提出了基于跳层卷积神经网络的RGB-D图像显著性检测方法。利用VGG网络分离出RGB图像和深度图像的浅层与深层特征,而后进行特征提取;以跳层结构为基础连接提取到的特征,实现融合深度、颜色、高级语义和细节信息的目标,同时生成侧输出;将侧输出进行融合,得到最佳的显著性检测图。实验结果表明,相比于深度监督显著性检测和渐进式互补感知融合显著性检测方法,在F值指标上分别提高了0.095 3和0.060 6,在平均绝对误差指标上降低了0.026 7和0.058 1。The saliency detection of RGB-D images aims to extract salient objects in three-dimensional images. In order to solve the problems of current saliency detection algorithms that are difficult to detect targets in light interference scenes and low-contrast targets, a saliency detection method for RGB-D images based on skip layer convolutional neural network is proposed. Firstly, the VGG network is adopted to separate the shallow and deep features of the RGB image and the depth image, and feature extraction is performed. Then the extracted features based on the skip-layer structure are connected to achieve the goal of fusing depth, color, high-level semantics and detailed information, and side outputs are generated.Finally, the side outputs are fused to obtain the best saliency detection map. Experimental results show that, compared with the deep supervised saliency detection and progressive complementary perception fusion saliency detection, the F value is improved by 0.095 3 and 0.060 6, and the average absolute error is reduced by 0.026 7 and 0.058 1, respectively.

关 键 词：显著性检测 卷积神经网络 跳层结构 深度学习 RGB-D 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]