毕业生知识库

工业外观质检是产品制造流程中的一个重要环节。基于深度学习的表面缺陷检测算法，由于其性能稳定且检测效率高等优点，在工业外观质检领域中得到了广泛应用。不过在较多的工业应用场景中，由于缺陷样本数量少、缺陷类别不均衡、样本标注困难等问题，导致难以构建理想的大规模缺陷数据集来训练有监督检测模型。图像异常检测算法可以仅使用正常样本构建模型，来定位图像中的各种异常区域，有助于推动深度学习在外观质检中的进一步应用。不过，由于工业图像与自然图像存在较大差异，导致一些通用图像异常检测算法难以直接应用在工业质检领域中。对于纹理背景各异的工业图像，如何实现异常区域定位成为了一个重要的研究课题。本文针对工业场景下图像异常区域定位的关键问题开展研究，以递进的方式逐步提升算法的适用范围与检测精度。本文的主要工作和贡献如下：

（1）针对产品表面纹理分布不均匀导致重构模型质量差的问题，本文提出了一种基于加权图像重构的异常定位方法。该方法在利用自编码器重构图像的基础上，设计了一种基于灰度方差加权的结构相似性损失函数，提升了模型在复杂纹理区域的重构精度。在重构差异分析阶段，利用基于离散余弦变换的区域分析模块来初步定位异常区域，在基本保证异常区域完整性的同时，去除纹理边缘的误检区域。实验表明本方法在保证实时性的同时，能够准确检测产品表面微小的异常区域。

（2）针对异常区域大小不一导致传统重构模型容易出现异常残留的问题，本文提出了一种基于异常样本映射学习的图像重构与异常定位方法。该方法首先将自编码器与生成式对抗网络结合来提升重构图像的清晰度。在重构模型训练过程中，添加了人造缺陷样本使网络学习特征空间中异常样本的映射过程，提升重构网络对大面积异常区域的重构质量。在完成图像重构后，采用一个分割模型来分析重构差异，在抑制正常区域重构差异对定位精度的影响的同时，实现像素级的异常区域定位。实验表明，该方法在多种复杂纹理表面上都能稳定获取无缺陷重构图像作为参考，提高了异常定位的精度。

（3）针对待测样本之间位姿多样导致难以用一个模型学习其数据分布的问题，本文提出了一种基于位姿匹配和距离度量的异常定位方法。该方法首先通过位姿匹配将输入图像与模板图像对齐，通过提高输入图像的一致性来降低模型的训练难度，便于后续的特征提取与异常定位过程。在距离度量阶段，为每个图像区域分配了一个局部特征映射中心，使正常图像的特征分布更加紧凑。此外，在模型训练阶段还合成多样化的人造缺陷图像，进一步提高异常定位模块的判别能力。实验结果表明，该方法在待测样本之间存在显著的位置与角度变化的情况下，取得了比现有方法更加优越的检测精度。

（4）针对异常检测精度要求高的问题，本文提出了一种基于缺陷预定位与注意力机制的异常定位方法。该方法在图像重构模型的基础上，利用缺陷预定位模块提前定位并消除输入图像中可能的异常区域，通过减少异常结构对重构的影响来提升图像重构的质量。随后，本文融合空间和通道注意力构建了混合注意力模块，来进一步优化图像重构的质量和分割模型的定位精度。此外，在模型的训练阶段还利用柏林噪声提升了人造缺陷样本的多样性。在开源图像异常检测数据集上的实验表明，本方法在与其余数种方法的性能对比中体现了优越性。

Industrial appearance quality inspection plays an important role in the product manufacturing process. Deep learning-based surface defect inspection methods are widely adopted in the industrial quality inspection field owing to their stable performance and high detection efficiency. However, due to the problems such as the class imbalance and the small number of defective samples, and the difficulties in annotating images, it is difficult to construct ideal large-scale defect datasets for training supervised detection models in many industrial production scenarios. Image anomaly detection, a method of training models using merely normal images to locate various anomalous regions in the test stage, may facilitate the further application of deep learning in appearance quality inspection. But the large differences between industrial and natural images limit the direct application of general image anomaly detection methods in the industrial quality inspection field. Designing anomaly localization methods for industrial images with diverse texture backgrounds has become an important research topic. This dissertation mainly focuses on several unresolved key issues of anomaly region localization in industrial scenarios, and gradually improves the applicability and accuracy of the method. The major works and contributions of this dissertation are summarized as follows:

(1) Aiming at the problem that the unevenly distributed textures of product surfaces lead to the poor quality of reconstructed images, this dissertation proposes an anomaly localization method based on weighted image reconstruction. In addition to the autoencoder, a loss function based on structural similarity weighted by gray variance is designed to improve the reconstruction quality of regions with complex textures. During the reconstruction difference analysis stage, a regional analysis module based on discrete cosine transform is utilized to initially locate the anomalous regions. The proposed module removes the false detection regions near the edges of textured while basically ensuring the integrity of the defects. It is demonstrated in experiments that this method can efficiently locate small anomalous regions on the surfaces of the products.

(2) Aiming at the problem that the traditional reconstruction models tend to leave some anomalous structures in the reconstructed images due to the varying size of defects, this dissertation proposes an image reconstruction method combining the learning of mapping method of anomalous images to locate defects. Firstly, the generative adversarial network is combined with the autoencoder to improve the clarity of the reconstructed images. Then, synthetic anomalous images are introduced in the training process of image reconstruction. Specifically, the encoder is prompted to learn the mapping method of the anomalous images in the feature space, improving the reconstruction quality of images with large anomalous regions. Following the image reconstruction, a difference analysis module based on image segmentation is adopted to locate anomalies pixel by pixel and suppress the influence of the noises in normal regions. Experiments conducted on various images with complex textures show that the proposed method can stably obtain defect-free reconstructed images as the reference, which improves the accuracy of anomaly localization.

(3) To reduce the difficulty of learning the distribution of samples to be tested with diverse postures, this dissertation proposes an anomaly localization method based on posture matching and distance measurement. Firstly, a posture matching module is designed to align the input images with a pre-specified template image. This method reduces the training difficulty of the model by improving the consistency of the input images, facilitating the subsequent feature extraction and anomaly localization processes. Afterward, each image region is assigned a local mapping center during the training stage of feature extraction, improving the compactness of representations of defect-free images. Furthermore, diverse artificial defective images are synthesized during the training stage to further improve the discriminant ability of the anomaly localization module. Experiments demonstrate that when spatial misalignment exists in the images to be tested, the proposed method achieves higher localization accuracy than the previous methods.

(4) To further improve the anomaly localization accuracy, this dissertation proposes a method combining defect perception and attention mechanism to precisely locate anomalous regions. Based on image reconstruction, this method adopts a defect perception module to locate and eliminate possible anomalous regions in the input images in advance. Therefore, this module can improve the quality of reconstructed images by reducing the impact of anomalous structures on the reconstruction. Then, spatial and channel attention is utilized to construct a hybrid attention module to further optimize the image reconstruction and anomaly localization processes. In addition, Berlin noise is adopted during the training stage to improve the diversity of artificial defective images. Experiments conducted on the open-source image anomaly detection dataset verify the effectiveness of the proposed method.