Abstract: A Weixian radish screening system based on external features has been designed to address current problem of manual selection of surface defects in Weixian radish, which results in high work intensity and low sorting efficiency．A test bench has been built to address this issue．Photos of surface defects in Weixian radish were collected on-site, and a dataset was constructed．Dataset was expanded by adjusting brightness and contrast, rotation, scaling, Gaussian blur, and adding salt and pepper noise to prevent overfitting．The data was annotated into four types: root multi head, bending, damage, and insect eye．YOLOv5s network with small model size and fast detection speed was selected to train Weixian radish surface defect dataset and generate a recognition model．Results showed that its recognition accuracy was 90.25%, accuracy was 93.77%, recall rate was 90.83%, and mAP@0.5 was 91.21%．Actual working environment was simulated, an experimental prototype was designed and produced, using a conveyor belt to drive radish to move．The upper computer collected and processed surface information of radish through dual cameras, and the lower computer controlled synchronous belt slide table to drive push plate to complete sorting of radish．Screening experiments were conducted at conveyor belt speeds of 0.16, 0.30, and 0.38 m/s on conveyor platform, and results showed that overall screening accuracy was 97.5%, 92.0%, and 82.5%, respectively．This study could provide design references for development of radish screening device in Weixian County．