Detection of Pine Cones in Natural Environment Using Improved YOLOv4 Deep Learning Algorithm

Achieving the rapid and accurate detection of pine cones in the natural environment is essential for yield estimation and automatic picking. However, the complex background and tiny target pose a significant challenge to pine cone detection. This paper proposes a pine cone detection method using the improved You Only Look Once (YOLO) version 4 algorithm to overcome these challenges. First, the original pine cone image data come from a natural pine forest. Crawler technology is utilized to collect more pine cone images from the Internet to expand the data set. Second, the densely connected convolution network (DenseNet) structure is introduced in YOLOv4 to improve feature reuse and network performance. In addition, the backbone network is pruned to reduce the computational complexity and keep the output dimension unchanged. Finally, for the problem of feature fusion at different scales, an improved neck network is designed using the scale-equalizing pyramid convolution (SEPC). The experimental results show that the improved YOLOv4 model is better than the original YOLOv4 network; the average values of precision, recall, and AP reach 96.1%, 90.1%, and 95.8%; the calculation amount of the model is reduced by 21.2%; the detection speed is fast enough to meet the real-time requirements. This research could serve as a technical reference for estimating yields and automating the picking of pine cones.