毕业生知识库

促进制造业转型升级与推动工业智能化发展是当今世界的发展趋势，对自动化缺陷检测技术的研究是其中重要的一环。异常检测是实现自动化缺陷检测算法的技术路线之一，与其他缺陷检测算法相比，异常检测算法的训练过程不需要缺陷数据及其标注，这使得异常检测算法具有成本低、适用范围广的优势。因此，对异常检测算法的研究可以为制造业转型升级与工业智能化发展提供技术支撑，具有重要现实意义。本文以基于图像重构的异常检测方法作为研究方向。并针对当前基于图像重构的异常检测算法中主要存在的异常漏检、异常过检、以及检测效率三个难点问题进行研究，提出了三个基于图像重构的异常检测算法。本文的主要工作和贡献在于：

（1）针对异常漏检问题，提出了基于参考图像辅助修复的异常检测算法。由于漏检问题是由输入图像中正常信息缺失过多导致的，因此该算法从向输入图像中补充正常图像信息的角度解决漏检问题。该算法首先引入训练集中与当前输入最相似的正常图像作为参考图像，并通过参考图像与输入图像之间的相似度对异常区域进行预定位。随后，该算法在预定位的异常区域生成混合图像掩膜，在保留输入图像信息的同时引入了参考图像上的正常图像信息。在该算法的重构网络前三层，本文设计了特征填充模块，特征填充模块通过区分掩膜区域特征与逐步修复掩膜区域的方式，进一步提升了模型对异常区域的修复能力，并增大异常区域在重构前后的差异，缓解漏检问题。

（2）同时针对异常漏检与异常过检问题，本文提出了融合先验信息的双支路异常检测算法。对于异常漏检问题，该算法直接从向重构网络中引入训练集均值图这一先验信息。考虑到不一定所有先验信息都是修复异常区域时需要的，本文设计了限制注意力机制用于筛选重构网络修复异常区域时需要的先验信息，并将筛选后的信息引入重构网络中。通过这种方式，该算法提升了重构网络对异常区域的修复能力，进而缓解异常漏检问题。
由于异常过检问题是由重构后的图像在正常区域的纹理等细节上与输入图像存在差异导致的，本文提出了双支路自编码器结构，该结构在重构正常图像外，也将输入图像的正常区域修改为与重构的正常图像一致。这使得以上两张图像只在异常区域存在差异，因此分割网络不会将输入图像中的正常区域判断为异常，进而避免异常过检问题。
 
（3）同时针对异常漏检、异常过检、以及检测效率问题，本文提出了融合重构与分割的异常检测算法。该算法首先将重构与分割融合在一个网络中，减少了一次图像特征提取过程，进而提升了模型的检测效率。由于该算法结构直接在特征图级别挖掘输入与重构之间的差异，充分利用了特征图信息，因此该算法结构也在一定程度上缓解了异常漏检问题。为了进一步解决异常漏检与过检问题，该算法在其分割模块中通过高层语义引导与多层语义监督两个模块解决漏检与过检问题。其中高层语义引导将高层特征图的语义信息引入低层特征图中，多层语义监督模块对分割模块中各层特征图上的语义信息进行监督与调整。以上两个模块的加入使得分割模块中各层特征图上的语义信息更为准确，因此可以同时缓解异常漏检与过检问题。

综上所述，本文提出了三种异常检测算法，缓解了当前基于图像重构的异常检测算法中存在的异常漏检、过检以及检测效率问题。本文的研究有利于提升异常检测算法的精度与效率，并为实现高效率、高准确性的自动化缺陷检测提供了有效的技术支撑。

Promoting the development of industrial intelligence is a important trend, and the research on automatic defect detection technology is an essential part of it. Anomaly detection is one of the approaches to realize automatic defect detection. Compared with other approaches, the defect data and annotation are not required for anomaly detection, which enables the anomaly detection to cost lower and suitable for more scenarios. Therefore, the research on anomaly detection is able to support the development of industrial intelligence. This thesis is based on the reconstruction-based anomaly approach. There are three difficult problems in the current reconstruction-based anomaly detection methods: anomaly escape problem, overkill problem, and detection efficiency problem. To solve the above problems, three reconstruction-based anomaly detection methods are proposed. The main contribution of this thesis are as follows:

(1) Aiming at the anomaly escape problem, an method named reconstruction by reference-assisted inpainting for anomaly detection is proposed, which solves the anomaly escape problem from the perspective of supplementing normal information to the input image. Firstly, the training image that is most similar to the current input is regarded as the reference image. And the abnormal region is pred-located by the similarity between the reference image and the input image. Then, this method generates a mixing image mask in the pre-located abnormal region, which introduces the normal image information on the reference image and reduces the influence of abnormal at the same time. In addition, this method designs a feature filling module, which is inserted to the first three layers of the reconstruction network. The feature filling module further improves the ability to repair anomaly by distinguishing the features of mask regions and gradually repairing the mask regions. The experiment shows that this method mitigates the anomaly escape problem effectively.

(2) Aiming at the anomaly escape and overkill problem, a method named dual branch learning with prior information is proposed. To solve the anomaly escape problem, this method regards the mathematical expectation map of the training data as the prior information, and introduces it into the reconstructive network. Considering that not all the prior information is effective to repair the abnormal regions, this method designs a gated attention unit to filtrate the useful prior information and introduces it into the reconstructive network. Therefore, the repair ability of the reconstructive network is enhanced, which enables this method to alleviate the anomaly escape problem.
Since the overkill problem is caused by the reconstruction error in the normal regions, a dual branch autoencoder structure is proposed. The dual branch autoencoder reconstructs a normal image, and maps the normal regions of the input image to be the same with the reconstructed normal image. Therefore, The input and the reconstructed normal image are only different in the abnormal regions. And the segmentation network will not judge the normal regions in the input image as anomalies, which enables this method to mitigate the overkill problem.

(3) Aiming at the anomaly escape, overkill and detection efficiency problem, a method combing the reconstruction and segmentation is proposed. The method integrates reconstruction and segmentation into a unified network, which reduces a feature extraction process and improves the detection efficiency. Since this method directly excavates the difference between input and reconstruction at the feature level, which makes full use of the information on the feature map, this method sometimes can alleviate the anomaly escape problem. To further solve the anomaly escape problem and overkill problem, two modules named high-level semantic guidance and multi-scale semantic supervision are proposed. The high-level semantic guidance module introduces the semantic information of the high-level feature map into the low-level feature map. And the multi-scale semantic supervision module adjusts the semantic information on the feature maps of each layer in the segmentation module. The above two modules enables the semantic information of each feature map in the segmentation module to be more accurate, which alleviates the anomaly escape problem and overkill problem at the same time.

In summary, three anomaly detection methods are proposed in this thesis, which finally mitigate the anomaly escape problem, overkill problem, and detection efficiency problem in the current reconstruction-based anomaly detection methods. This thesis is conducive to improving the accuracy and efficiency of anomaly detection algorithm, and also provides effective technical support for realizing automatic defect detection with high efficiency and accuracy.