22 November 2021 in IEEE Transactions on Instrumentation and Measurement
Deep learning based algorithms have been widely employed to build reliable steel surface defect detection systems, which are important for manufacturing. The performance of deep learning models relies heavily on abundant annotated data. Nevertheless, the labeled image volume in industrial datasets is often limited. The scarcity of training data would lead to poor detection precision. To tackle this issue, we propose the first few-shot defect detection framework. Through pre-training models using data relevant to the target task, the proposed framework can produce well-trained networks with a few labeled images. Meanwhile, we release the first publicly available few-shot defect detection dataset, namely few-shot NEU-DET (FS-ND). This dataset will serve as a fair benchmark for contrasting various methods. Afterwards, we analyze the characteristics of steel surface defect detection. It is observed that the limited amount of training data can hardly cover the data distributions in practical applications. Given this observation, we develop two domain generalization strategies that enhance the appearance and scale diversity of extracted features. Furthermore, it is found that noise existing in industrial images could result in the collapse of models. To address this problem, we devise a noise regularization strategy that improves the robustness of trained models significantly. We have conducted extensive experiments to evaluate the effectiveness of our framework. The results indicate that our framework outperforms the contrasted baseline by around 15 mAP and achieves comparable performance with models trained using abundant data.
26 April 2022 in IEEE Transactions on Artificial Intelligence
Few-shot classification is challenging due to the limited data and labels. Existing algorithms usually resolve this problem by pre-training models with a considerable amount of annotated data which shares knowledge with the target domain. Nevertheless, large quantities of homogenous data samples are not always available. To tackle this obstacle, we develop a few-shot learning framework that prepares data automatically and still produces well-behaved models. This framework is implemented through conducting contrastive learning on unlabeled web images. Instead of requiring manually annotated data, this framework trains models via constructing pseudo labels. Additionally, since online data is virtually limitless and continues to be generated, the model can thus be empowered to constantly obtain up-to-date knowledge from the Internet. Furthermore, we observe that the generalization ability of learned representation is crucial for self-supervised learning. To present its importance, a naive yet efficient normalization strategy is proposed. Consequentially, this strategy boosts the accuracy of trained models significantly. We demonstrate the superiority of the proposed framework with experiments on miniImageNet, tieredImageNet and Omniglot. The results indicate that our method has surpassed previous unsupervised counterparts by a large margin and obtained performance comparable with some supervised ones.