Improved Method for Oriented Waste Detection

Waste detection is one of the main problems preventing realization automated waste classification, which a basic function for robotic arms. In addition to object identification in general image analysis, waste-sorting arm not only needs identify target but also accurately judge its placement angle so that it can determine an appropriate grasping. order solve problem low-accuracy caused by irregular angles, this work, we propose improved oriented method based on YOLOv5. By optimizing head YOLOv5 model, generate box placed at any angle. Based proposed scheme, further three aspects performance objects: angular loss was derived dynamic smoothing enhance model’s prediction ability, backbone network optimized with enhanced shallow features and attention mechanisms, feature aggregation effects multi-scale fusion. The experimental results showed targets better than other deep learning methods. Its average accuracy recall were 93.9% 94.8%, respectively, 11.6% 7.6% higher those original network, respectively.