目的 基于深度学习的工业缺陷检测方法可以降低传统人工质检的成本, 提升检测的准确性与效率，因而在智能制造中扮演重要角色。针对无监督工业缺陷检测中存在的过检测和逻辑缺陷检测失效等问题，提出一种融合知识蒸馏与记忆机制的无监督工业缺陷检测模型。方法 使用显著性检测网络和柏林噪声合成缺陷图像，提升合成图像与真实缺陷图像的分布一致性，缓解传统模型的过检测问题；同时，对传统无监督工业缺陷检测框架进行改进，引入平均记忆模块提取正常样本的原型特征，通过记忆引导提高模型对逻辑缺陷的检测性能。结果 在工业缺陷检测基准数据集MVTec AD上的实验表明，针对晶体管逻辑缺陷检测难题，在像素级接受者操作特征曲线下面积(area under the receiver operating characteristic curve, AUROC)指标上所提出的方法相比于基线模型提升了9.1%；针对各类缺陷检测场景，在更具挑战性的平均准确率(average precision, AP)指标上提升了2.5%。针对更具挑战性的早餐盒数据集中的逻辑缺陷问题，所提出的方法在图像级AUROC指标上相较于基线模型提升了11.5%。同时，在像素级AUROC指标上，所提出的方法相较于基线模型提升了4.0%。结论 所提出的方法不受传统缺陷合成方法的限制，能够有效缓解现有缺陷合成方法引起的过检测问题；引入平均记忆模块不仅可以减小内存开销，而且无需设计复杂的检索算法，节省了检索内存库所耗费的时间；将所提出的缺陷合成方法与记忆机制进行有机结合，能够准确地检测出不同种类的工业缺陷。

Objective From airplane wings to chip grains, industrial products are ubiquitous in modern society. Industrial defect detection, which aims to find the appearance defects of various industrial products, is one of the important technologies to ensure product quality and maintain stable production. Previous defect detection requires manual screening, high cost, low efficiency, difficult to cover large-scale quality inspection needs. In recent years, with the endless emergence of new technologies in the fields of industrial imaging, computer vision and deep learning, vision-based industrial defect detection technology has been greatly developed, and has become an effective solution for product appearance quality inspection. However, there are many types of industrial defects in the actual scene and there is a lack of samples, so the existing unsupervised industrial defect detection methods are difficult to effectively detect the local normal logic defects, such as the normal target appears in the wrong position or the target is missing. This is due to the lack of prior knowledge of the normal sample during the testing phase, resulting in the defective part being incorrectly identified as normal. In addition, deep neural networks have a strong generalization ability, and existing methods often detect interference factors on the background of the image as defects, resulting in over detection problems. Aiming at the problems of logic defect detection failure and over detection in unsupervised industrial defect detection, a new unsupervised industrial defect detection model is proposed. Method Firstly, a saliency detection network and Berlin noise are used to synthesize defect images, improving the distribution consistency between synthesized images and real defect images, and alleviating the problem of over detection in traditional models; Secondly, our model consists of the teacher-student branch and the memory branch. The teacher-student branch trains the student network by distilling knowledge and synthesizing defect images, so that it cannot only extract the normal image features consistent with the teacher network, but also repair the defect part, effectively alleviating the overgeneralization problem of the student network. By introducing the average memory module, the memory branch can effectively learn the prototype features representing normal samples and enhance the ability of the model to detect logical defects. The two branches adaptively fuse multi-scale defect features, and realize the accurate detection of various defects through joint discrimination. Result The experiments on MVTec AD, a benchmark data set for industrial defect detection, show that the proposed method achieves good detection performance for all kinds of defect images. For texture defect images, compared with the baseline model DeSTSeg, the average image-level AUROC metric was further increased from 99.3% to 99.8%, and the average pixel-level AUROC metric was further increased from 98.1% to 98.7%. For object class defect images, compared with the baseline model DeSTSeg, the average image-level AUROC metric was further increased from 97.5% to 99.1%, and the average pixel-level AUROC metric was further increased from 97.9% to 99.1%. Especially for the problem of transistor logic defect detection, the proposed method was improved by 9.1%. In the whole MVTec AD dataset, compared with the baseline model DeSTSeg, the average image-level AUROC metric was further increased from 98.1% to 99.3%, and the average pixel-level AUROC metric was further increased from 97.9% to 98.9%. In addition, the proposed approach achieved a 0.9% and 2.5% improvement on the more challenging pixel-level PRO (per-region-overlap) and average accuracy AP metrics, respectively. Addressing logic defects in the more challenging breakfast box dataset, the proposed method achieved an 11.5% improvement in image-level AUROC metrics compared to the baseline model. At the same time, in terms of pixel-level AUROC index, the proposed method improves by 4.0% compared with the baseline model. In the ablation experiment, each module of the proposed method is validated. The introduction of significance detection to constrain synthetic defects in the foreground can significantly reduce the over detection phenomenon caused by background interference of the model, and its classification performance is improved by 1% compared with the baseline model. With the addition of memory branches, the model can effectively detect logic defects and significantly improve segmentation performance. However, the direct average fusion method will damage the respective advantages of the two branches, resulting in poor defect detection performance. For this purpose, the use of the normalization module effectively combines the advantages of the two segmentation networks, which improves the classification and segmentation performance by 0.7% and 0.5%, respectively, compared to the direct averaging approach. Conclusion The proposed method is not limited by traditional defect synthesis methods, and can effectively alleviate the over detection problems caused by existing defect synthesis methods. The introduction of average memory module can not only reduce the memory cost, but also save the time of searching memory library without designing complicated search algorithm. In this paper, the proposed defect synthesis method is organically combined with the memory mechanism, which can accurately detect different kinds of industrial defects.