EN-YOLO: a featured-enhanced model for precise hot rolled steel surface defect detection

In industrial applications, the accurate detection of defects on hot-rolled steel surfaces is crucial, as even minor imperfections can lead to potentially hazardous situations. To address the need for precise defect detection, this study introduces a feature-enhanced YOLOv7 network model, termed EN-YOLO. This model is optimised using the EfficientNet-Block (EN-Block) module, specifically designed to enhance feature extraction and integrate contextual information by balancing network depth, width, and resolution. In addition, an attention mechanism is incorporated to prioritise critical areas in the analysis. Moreover, the contextual information integration is facilitated by the Bi-directional Feature Pyramid Network module, which enables interaction and communication of features across different levels. Furthermore, the introduction of the Residual Maxpooling module aids the network model in learning finer features, reducing information loss, and maintaining a smooth flow of information. As a result of these enhancements, the EN-YOLO model achieves a notable 82.7% mAP@0.5 on the NEU-DET dataset, representing a 4.5% increase in accuracy compared to the traditional YOLOv7 model, demonstrating the effectiveness of our proposed model and optimisation approach. Further validation on the public ASSDD and WSDD datasets confirms the superior generalisation performance of the EN-YOLO model.