Surface defect recognition is important to improve the surface quality of end products. In this area, there were many convolutional neural network (CNN)-based methods because CNN can extract features automatically. The extracted features determine the performance of recognition, so it is important for CNN-based methods to extract effective and sufficient features. However, feature extraction needs a large-scale dataset, which is hard to obtain. To save the cost of collecting samples and extract effective features, ensemble methods were proposed to make full use of the features extracted by CNN in order to guarantee good performance with limited samples. However, the methods are confined to utilize one sample – they extracted multi-level features from one individual sample – but ignore the vast information in a dataset. Due to the limit information in one sample, this paper turns the attention to the training dataset and attempts to mine the multi-level information in the dataset for predicting. The proposed method is named as Prototype vectors fusion-based CNN (ProtoCNN), which utilizes the prototype information in the training dataset. In training process, it trains a VGG11 as the base model, and meanwhile prototype vectors corresponding to each defect class are generated in multiple feature layers of VGG11. Then, in predicting process, the prototype vectors are fused to predict unknown samples. The experiments on three famous datasets, including NEU-CLS, wood dataset, and textile dataset indicate that the proposed ProtoCNN outperforms conventional ensemble models and other models for surface defect recognition. In these datasets, ProtoCNN has achieved the accuracy of 99.86%, 90.01%, and 81.28% respectively, which increase 1.05%, 4.07%, 19.53% compared to its base model respectively. Finally, this paper analyzes the effectiveness and practicality of prototype vectors, showing that the proposed ProtoCNN is practical for real world application.

表面缺陷识别对于提高最终产品的表面质量非常重要。在这方面，有很多基于卷积神经网络（CNN）的方法，因为 CNN 可以自动提取特征。提取的特征决定了识别的性能，因此基于CNN的方法提取有效且充分的特征非常重要。然而，特征提取需要大规模的数据集，很难获得。为了节省采集样本的成本并提取有效特征，提出了集成方法来充分利用CNN提取的特征，以保证有限样本的良好性能。然而，这些方法仅限于使用一个样本——它们从一个单独的样本中提取多级特征——而忽略了数据集中的大量信息。由于单个样本的信息有限，本文将注意力转向训练数据集，尝试挖掘数据集中的多层次信息进行预测。所提出的方法被命名为基于原型向量融合的CNN（ProtoCNN），它利用了训练数据集中的原型信息。在训练过程中，训练一个VGG11作为基础模型，同时在VGG11的多个特征层中生成每个缺陷类别对应的原型向量。然后，在预测过程中，融合原型向量来预测未知样本。在三个著名数据集（包括 NEU-CLS、木材数据集和纺织品数据集）上的实验表明，所提出的 ProtoCNN 优于传统的集成模型和其他表面缺陷识别模型。在这些数据集中，ProtoCNN 的准确率分别达到了 99.86%、90.01% 和 81.28%，相比其基础模型分别提高了 1.05%、4.07%、19.53%。最后，本文分析了原型向量的有效性和实用性，表明所提出的 ProtoCNN 在实际应用中是实用的。