ESMNet: An enhanced YOLOv7-based approach to detect surface defects in precision metal workpieces.

Surface defect detection in precision metal workpieces is critical for ensuring product quality. Due to the weak and diverse defect object area in precision metal workpieces, they lead to high missed and false inspections. This paper proposes a novel machine vision solution to address challenges, focusing on two aspects. First, we develop a raster rotation imaging method, which shows defect area more salient than general imaging. Second, we propose a modified detection model named ESMNet based on YOLOv7-tiny by integrating the ELAN-SC module, which is designed to reduce feature redundancy and enhance salient feature learning, and the Multi-Scale Cross Fusion Attention (MCF) module, which helps to aggregate local contexts of weak defect area. Additionally, we gather defect samples to construct the CSD-DET dataset of defects on cylindrical surface to support the training and evaluation of detection models. Experiments on the CSD-DET dataset demonstrate that ESMNet achieves a 0.8% improvement in mean Average Precision (mAP) over state-of-the-art methods while maintaining low computational complexity, showcasing outstanding detection performance. This solution is applicable not only to cylindrical surface of metal workpieces but also to promote the application of machine vision in complex industrial environments. • Introduces a novel method for detecting defects on the cylindrical surfaces of precision metal workpieces. • The proposed raster rotation imaging makes defects more salient. • Constructs CSD-DET dataset using samples collected from workshops for training and validating. • The proposed ESMNet demonstrates excellent detection performance. [ABSTRACT FROM AUTHOR]