Design and implementation of a deep learning-based quality detection system for cork discs

High-quality badminton shuttlecock heads are made from cork discs derived from natural oak bark, which exhibit complex cracks and wrinkles, causing inefficiency and inconsistency in quality screening. To address this, we propose a deep learning-based detection algorithm using YOLOv5. The YOLOv5 model was optimized for cork disc characteristics by introducing an attention mechanism to enhance feature representation and designing a post-processing algorithm to improve detection accuracy. The optimized model, trained on a custom cork disc dataset using an NVIDIA RTX3080 GPU, achieved 86.7% mF1 and 81.5% mAP, outperforming other mainstream algorithms. Finally, the system utilizes Nvidia's edge computing device Jetson Nano as the computational core, deploying the YOLOv5 model and designing the graphical interface on Ubuntu 18.04. Real-time cork disc image acquisition is achieved using binocular industrial cameras and fiber optic sensors, while a uniformly rotating turntable and diversion device are designed to facilitate the transfer of cork disc. Experimental results show that the system achieves a 92.75% classification accuracy for four quality grades with an inference time of 87.6ms, meeting the requirements for real-time cork disc quality inspection.