【作者】 王健；

【导师】 陈文兵；

【作者基本信息】 南京信息工程大学 ， 数学， 2023， 硕士

【摘要】 显著目标检测旨在模拟人的视觉系统,检测并分割出最吸引人注意的目标和图像区域。在许多计算机视觉任务中,显著目标作为预处理工作发挥着重要的作用。由于深度图包含RGB图像所没有的空间结构信息,近年来随着深度传感器的发展及获取深度图越来越便捷,相关领域越来越多的研究人员将研究兴趣投向基于RGB-D的显著目标检测。迄今已取得许多引人注目的研究成果,但仍然存在许多亟待解决的问题。本文针对现有的问题,提出了两个RGB-D显著目标检测模型。(1)大多数RGB-D显著目标检测模型利用深度图与RGB图相互补充来检测显著目标。然而,由于RGB-D显著目标检测数据集中存在大量低质量深度图,这给模型最终的显著检测带来消极影响。为解决这一问题,提出一种三数据流互补的显著目标检测模型。首先,挑选优质深度图作为RGB数据流的训练目标,将其对应的RGB图作为该数据流的输入进行训练,从而所训练的RGB数据流能够为每幅RGB图生成一幅深度估计图。接着,利用一个改进的特征互补融合模块来融合原始深度图、深度估计图及RGB图三个数据流显著特征。最后,使用具有大尺度感受野的解码器对融合的三数据流特征进行自顶向下地解码以获得最终的目标显著图。与10种主流模型在4个基准数据集上的对比实验表明,所提模型在4个评价指标上均有明显提升。(2)现有的RGB-D显着目标检测模型大多采用VGGs和ResNets等大型网络作为主干,这导致模型规模过大、计算成本过高。为了改善这一问题,提出一个轻量的两阶段解码的显著目标检测模型。首先,利用Mobile Net-V2和定制的骨干网分别提取RGB图像和深度图的特征。其次,为了挖掘和组合跨模态信息,使用交叉引用模块来融合两种模态中的互补信息。然后,使用一个特征精炼模块来增强融合特征的信息。最后,采用一个两阶段解码器分别整合高阶特征和低阶特征来获取最终的显著预测图。与10个主流模型在5个基准数据集上的对比实验表明,所提模型计算成本明显降低。更多还原

【Abstract】 Salient Object Detection aims to simulate the human visual system to detect and segment the most attractive targets and image regions.In many computer vision tasks,salient objects play an important role as a preprocessing step.Since depth maps contain more spatial structure which is not available in RGB images,recently with depth sensors developed and depth maps more easily obtained,more and more researchers in relevant fields are interested in SOD based on RGB-D.So far many remarkable research results have been achieved,while there still exist many issues to urgently be addressed.In this paper,two RGB-D salient object detection models are proposed to address the existing problems.(1)Most existing RGB-D salient object detection models use depth maps and RGB maps to complement each other for salient object detection.However,there are numerous low-quality depth maps in RGB-D salient object detection datasets,which affect the final saliency prediction negatively.In order to deal with this problem,a novel three-stream complementary network is proposed.Firstly,high-quality depth maps are chosen as training targets of RGB stream,and the corresponding RGB maps are used as inputs for training.In this way,the trained RGB stream can generate an estimated depth map for each RGB map.Secondly,a feature complementation fusion module is constructed to merge original depth stream features,estimated depth stream features and RGB stream features.Finally,a top-down decoder with large-scale receptive field is designed to decode salient features from different stages and predict saliency maps.Experiments on 4 benchmark datasets comparing with 7 state-of-the-art models demonstrate that our model can achieve obvious improvement.(2)Most existing models of RGB-D salient object detection utilize heavy backbones like VGGs and Res Nets which cause large model size and high computational costs.In order to address this problem,a lightweight two-stage decoder network is proposed.Firstly,Mobile NetV2 and a customized backbone are used in the network to extract the features of RGB images and depth maps respectively.In order to mine and combine cross-modality information,cross reference module is used to fuse complementary information from different modalities.Subsequently,a feature enhancement module is designed to enhance the clues of the fused features which has four parallel convolutions with different expansion rates.Finally,a two-stage decoder is used to predict the saliency maps,which processes high-level features and low-level features separately and then merges them.Experiments on 5 benchmark datasets comparing with 10 state-of-the-art models demonstrate that our model can achieve obvious improvement with smallest model size.更多还原