Apple Detection in Complex Scene Using the Improved YOLOv4 Model

To enable the apple picking robot to quickly and accurately detect apples under complex background in orchards, we propose an improved You Only Look Once version 4 (YOLOv4) model data augmentation methods. Firstly, crawler technology is utilized collect pertinent images from Internet for labeling. For problem of insufficient image caused by random occlusion between leaves, addition traditional techniques, a leaf illustration method proposed this paper accomplish augmentation. Secondly, due large size calculation YOLOv4 model, backbone network Cross Stage Partial Darknet53 (CSPDarknet53) replaced EfficientNet, convolution layer (Conv2D) added three outputs further adjust extract features, which make lighter reduce computational complexity. Finally, detection experiment performed on 2670 expanded samples. The test results show that EfficientNet-B0-YOLOv4 has better performance than YOLOv3, YOLOv4, Faster R-CNN with ResNet, are state-of-the-art model. average values Recall, Precision, F1 reach 97.43%, 95.52%, 96.54% respectively, time per frame 0.338 s, proves can be well applied vision system robots industry.