【作者】 李振杰；

【导师】 姜晓恒； 徐明亮；

【作者基本信息】 郑州大学 ， 计算机技术（专业学位）， 2024， 硕士

【摘要】 表面缺陷检测是在产品生产中保证产品的质量、提高生产效率的一个重要环节。随着科技的发展以及对产品质量要求的提高,基于深度学习的表面缺陷检测方法取代了人工检测方法并被广泛的研究。目前使用可见光图像进行表面缺陷检测取得了巨大的发展,但当遇到缺陷表面颜色和纹理与正常表面高度相似时,仅靠可见光图像无法精确的检测出缺陷区域。深度图像中具有大量的空间信息,图像展示出的缺陷区域与正常表面区域会有明显的差异,可以与可见光图像信息互补提升缺陷检测精度。所以如何有效的挖掘出RGB-D(RGB-Depth)之间的互补信息,充分的融合RGB-D特征,实现准确的表面缺陷检测,具有很大的研究意义。目前基于深度学习的RGB-D表面缺陷检测方法还存在一些挑战,例如RGB-D各模态之间特征无法充分融合、模态之间存在差异性和深度图像质量参差不齐,针对这些问题,我们开展了一系列的研究,主要贡献如下:1.针对当前方法RGB-D图像特征融合不充分的问题,本文提出了基于RGB-D的双模态注意力交互融合的表面缺陷检测方法。该方法设计了双模态注意力交互融合模块,结合通道注意力和空间注意力机制,实现跨模态的注意力交互融合,使得网络更好的学习双模态的共有特征以及特有特征并进行充分的融合。并采用多尺度空间注意力方法,学习跨模态的全局关系,通过区域对齐解决模态间存在差异性的问题,减少因为多模态差异带来的噪声,提高缺陷检测网络检测的性能。在NEU RSDDS-AUG缺陷数据集上证明了该方法在表面缺陷上的有效性,在NLPR和NJU2K数据集上的验证了缺陷检测网络的泛化能力。2.为了解决深度图像质量低时对缺陷检测网络带来的负面影响,本文提出了一种基于自适应深度质量评价的RGB-D表面缺陷检测算法。该方法使用了一种级联分层的方法,对深度图像质量进行评估,通过非局部注意力来学习模态的全局依赖关系,分析两种模态间的差异性,产生深度图像贡献权重,控制深度图像在融合时的贡献,深度图像质量越差权重越小,进而减少错误信息对缺陷检测网络的影响,提高缺陷检测网络的鲁棒性。在NEU RSDDS-AUG缺陷数据集上表明了方法在表面缺陷上的有效性和鲁棒性。更多还原

【Abstract】 Surface defect detection is an essential aspect in ensuring product quality and enhancing production efficiency in product manufacturing.With the development of technology and increasing demands for product quality,deep learning-based surface defect detection methods have replaced manual inspection methods and are widely researched.Currently,using visible light images for surface defect detection has made significant progress.However,when faced with surface defects that have similar colors and textures to normal surfaces,it is challenging to accurately detect the defective areas solely based on visible light images.Depth images contain abundant spatial information,and the differences between defective and normal surface areas in these images are significant,complementing the information from visible light images and enhancing defect detection accuracy.Therefore,effectively mining the complementary information between RGB-D(RGB-Depth)and fully integrating RGB-D features for accurate surface defect detection is of great research significance.This paper addresses the problem of insufficient integration of features between RGB-D modalities,the differences between modalities,and the inconsistent quality of depth images,and conducts a series of studies,with the main contributions as follows:1.In response to the current inability to sufficiently integrate RGB-D image features,this paper proposes a surface defect detection method based on RGB-D with dual-modal attentional interaction fusion.A dual-modal attentional interaction fusion module is designed,which combines channel attention and spatial attention mechanisms to achieve cross-modal attentional interaction fusion,enabling the network to better learn both shared and unique features of the dual modalities and integrate them fully.Additionally,a multi-scale spatial attention method is employed to learn global relationships across modalities,addressing the differences between modalities through region alignment to reduce noise caused by multimodal discrepancies and improve the performance of the defect detection network.The effectiveness of this method in surface defect detection is demonstrated on the NEU RSDDS-AUG defect dataset,and its generalization capability is validated on the NLPR and NJU2K datasets for defect detection.2.To address the negative impact on defect detection networks caused by lowquality depth images,this paper proposes an RGB-D surface defect detection algorithm based on adaptive depth quality assessment The method employs a cascaded hierarchical approach to evaluate the quality of depth images.It utilizes non-local attention to learn global dependency relationships among modalities,analyzes the differences between the two modalities,generates depth image contribution weights,and controls the contribution of depth images during fusion.The contribution weight decreases as the depth image quality worsens,thereby reducing the influence of erroneous information on defect detection network and enhancing its robustness.The effectiveness and robustness of the proposed method in surface defect detection are demonstrated on the NEU RSDDS-AUG defect dataset.更多还原