The current chopstick quality inspection machines on the market cannot effectively sort chopped chopsticks with burrs. Aiming at this problem, in this paper, a method for detecting burr defects of chopsticks based on improved YOLOv3 algorithm is proposed. By removing the 32x down-sampling detection layer in the YOLOv3 network multi-scale detection, adding a 4x down-sampling layer in the YOLOv3 network to further obtain deep features. Thereafter, it was fused with the shallow features in the second down-sampling, and let the network learn the deep and shallow features and re-cluster the anchor box size, with changing the hyper-parameters of the YOLOv3 network, including reducing jitter and the weight-decay regular term, and increasing the batch size. Finally, a suitable momentum value was selected to improve the original network. When IOU=50, the average detection accuracy of the improved network increased from 89% to 94%, and the accuracy rate increased by 4%, with the recall rate increasing by 9% and the average IOU increasing by 3.5%. The average detection speed increased from 16.8 to 21.0 frames per second. The experimental results showed that the method in this study had higher detection efficiency than the traditional chopstick quality inspection machine, which could meet the detection needs of chopstick burr defects.