Detection of Apple Surface Defect Based on YOLOv3

Abstract. In order to realize the rapid and accurate identification of apple defects, apple surface defects were identified by the YOLOv3 target detection model in this paper. The model included a residual structure, which increased the depth while ensuring the convergence of the model. A multi-scale feature extractor was used for target detection in this model, which improved the detection effect but didn‘t increase the amount of parameters, ensuring the detection speed. After training and comparison, 0.50 was selected as the confidence threshold for apple defect detection. The Precision of the detection result was 97%, the Recall was 87%, and the comprehensive evaluation F1 was 0.92. Compared with the other three target detection models of SVM, Faster RCNN, and YOLOv2, F1 had increased by 5.6%, 3.64%, and 1.45% respectively. In addition, the average detection time of YOLOv3 for a single image was 1.12s, which was faster than the other three models. In order to verify the actual detection speed, the model was tested on the CPU and GPU with the same data set. The results showed that the detection speed of the CPU was significantly slower than the detection speed of the GPU. In this test, the average detection time of a single image of the CPU exceeded 1s, which failed to meet the real-time requirements, while the average detection time of a single image of the GPU was 42ms, which met the real-time requirements.