An efficient and lightweight rotating target detection model for industrial scenarios

Abstract

Currently, target detection technology has largely matured in industrial scenarios, but most applications still rely on horizontal bounding boxes for target detection. For industrial parts with irregular shapes, varying sizes, and diverse categories, horizontal bounding box detection tends to introduce unnecessary background information, leading to false positives or missed detections. It also suffers from the issue of losing boundary direction. To address these challenges, this paper proposes a novel YOLO11 model for detecting rotating targets (DAGP-YOLO). First, a rotating detection head is introduced to minimize the interference of redundant background information. Dynamic convolution is applied to expand the receptive field of the network, and the ADown module replaces the original down sampling method to improve detail extraction. Second, an orientation-aware attention mechanism (GCA) is designed to better focus on the directional features of rotating targets. Lastly, to meet the demand of small storage space and high detection accuracy of edge devices in industry, this paper adopts the L1 filter pruning strategy to compress the improved model. We performed experimental validation on a self-constructed dataset, the publicly available MVTec Screws dataset, and the UCAS AOD dataset in the aerial photography domain. The results demonstrate the superiority and effectiveness of our approach. Additionally, we developed a visualization system based on the C# WinForms framework, which allows for real-time detection and display of workpiece images, further showcasing the practical applicability of the improved method in industrial settings.