Online visual end-to-end detection monitoring on surface defect of aluminum strip under the industrial few-shot condition

Surface defect detection systems based on deep learning are employed in the manufacturing system, and their good detection performance largely relies on abundant annotated data. Nevertheless, industrial datasets are often difficult to obtain, which hinders the development of defect detection systems to some extent. In an effort to address the issue that the scarcity of training data leads to poor performance, this paper proposes a novel end to end few-shot detection method for industrial real-time detection on aluminum strips. The meta-learning theory is introduced into the multi-scale structure of the YOLOv4 framework, which means that multi-scale meta-feature maps are extracted through the backbone network and each meta-feature map is recalibrated by reweighted vectors to strengthen class-specific features. Subsequently, a multi-scale prediction module with a new parsing strategy is designed to locate and classify the defects at different scales for the reweighted meta-features. All experiments are performed using data collected from the aluminum strip production line in a cold rolling workshop. In addition, the dataset required for training is constructed based on the few-shot training strategy. Pre-training is conducted on a large amount of labeled data from base classes in advance, and the pre-trained weights are then loaded to train the model on a small amount of labeled data from all classes to rapidly deploy the model in multi-task scenarios. The proposed method shows excellent performance, improving the accuracy for novel classes by more than 14.6%, which enhances the industrial practicality of detection systems.