目的 多类缺陷检测是工业检测领域中的重要应用场景，现有方法因需训练多个模型而导致其耗时耗内存，同时多类之间因存在特征扰动导致现有模型鲁棒性欠佳。为此，本文联合特征扰动池和多层特征融合提出一种多类缺陷检测的新网络。方法 一方面通过特征扰动池强化模型对特征的鲁棒性，另一方面将各层网络输出特征进行融合，从而降低模型对数据分布的依赖性和提升对特征间复杂关系的捕捉能力。结果 与SOTA方法相比，所提方法在缺陷检测和缺陷定位方面具有出色的性能，在MVTec-AD数据集上分别达到了97.1%和96.9%，而在VisA数据集上也分别达到了91.0%和99.0%。结论 提出的联合特征扰动池和多层特征融合的多类缺陷检测网络具有更好的鲁棒性，能够捕捉特征之间的复杂关系，可广泛应用于工业缺陷检测领域。

Objective In the industrial manufacturing domain, quality control is the key link to ensure the reliability degree and safety of products. As a key part of quality control, the efficiency and accuracy of defects detection directly affect product quality and the economic benefits of enterprises. Traditional defects detection methods, such as artificial vision detection or rule-based image processing techniques, are often limited by the subjective judgment of operators and the complexity of rules, and are difficult to adapt to the needs of modern industry for high efficiency and high accuracy. With the development of deep learning technology, especially the breakthrough of deep neural network in the image identification domain, defects detection technology based on deep learning has gradually become a research hotspot. Current defects detection solutions are usually modeled based on the data distribution of normal samples and detect and label defect samples by identifying outliers in the data distribution. However, in practical applications, there are many defect types of products with different defect morphologies, and it is difficult for a single model to effectively cover all defect types. Therefore, it has become a common practice to train a separate model for each defect type. However, this practice not only requires a lot of training data and computing resources, but also the maintenance and update of the model is extremely cumbersome. To solve these problems, researchers have designed a unified detection framework that can handle multiple defect types in a single model. However, these unified frameworks are often trained with a single fixed perturbation scheme, which limits the model's ability to learn various flawed features. In addition, these frameworks fail to fully utilize the feature information output by encoders and decoders at various layers in the network, resulting in models that may over-rely on specific sampling features, thereby reducing the generalization ability of the model. Method Aiming at the problems existing in the unified detection framework, this paper proposes an innovative improved multi-class defects detection network, which significantly improves the generalization ability and robustness of the model by introducing a feature perturbation pool and a multi-layer feature fusion strategy. On the one hand, in traditional deep learning model training, the model is usually trained only on raw datasets, which causes the model to be over-fitting to a specific data set, thus performing poorly in the face of various changes that may be encountered in actual production. To overcome this problem, the feature perturbation pool adds a series of stochastic perturbations to the training data during training. These perturbations can be rotation, scaling, cropping, color dithering, etc. of the input data, simulating various changes that may be encountered in actual production. Through such data augmentation, the model is able to learn a more generalization of feature representations, which makes it more adaptable in practical applications. On the other hand, in traditional deep learning model training, usually only the features of the last layer are used for the final classification or task. However, feature maps at different levels contain feature information at different scales and levels of abstraction, which is crucial for defect identification and localization. Therefore, this paper proposes a multi-layer feature fusion strategy that integrates feature information at different levels in the network. Specifically, we use the skip connection method to directly fuse the low-level feature map in the encoder with the high-level feature map in the decoder. In this way, the model can not only capture global context information, but also retain local details, thus improving the ability to identify and locate defects. Result Compared with the current state-of-the-art methods, our proposed method exhibits excellent performance. The defect detection accuracy and the defect localization one reach 97.1% and 96.9% on MVTec-AD dataset, respectively. And these two types of accuracy also reach 91.0% and 99.0% on VisA dataset, respectively. Conclusion The joint feature perturbation pool and multi-layer feature fusion multi-type defects detection network proposed by us have better robustness and can capture the complex relationship between features. The network not only shows great potential in theoretical exploration, but also shows broad application prospects in practical industrial applications. In the future development, exploring more efficient feature extraction and fusion technologies and smarter training strategies to deal with more diverse defect types and more complex production environments can promote the development of industrial quality control systems to a smarter and more efficient direction, making greater contributions to improving product quality and production efficiency.