关键词: 经济林木, 虫害, YOLOv5, 深度学习, 特征融合, 目标检测

Abstract: Object detection of insect pests in economic forests helps to detect insects in a timely manner, so as to control the pests more pertinently. In this paper, the weighted bidirectional feature fusion technology was used to enrich the semantic information of feature maps at all levels and the adaptive Anchor calculation method was modified to improve the YOLOv5 backbone network model. Then experiments were carried out on the enhanced data set containing 20 kinds of economic forest pests, and finally compared with a variety of object detection methods based on deep learning in map and inference speed. The results showed that compared with YOLOv3, YOLOv4, YOLOv5, FasterRCNN and CenterNet models, the average accuracy of the improved YOLOv5 model was increased by 0.133、0.156、0.113、0.128 and 0.078, respectively, with the optimal accuracy of 0.923, and the model inference speed was 64.9 frames. Therefore, the improved YOLOv5 model in this paper has reached the practical application level on detecting insect pests in economic forests, which can provide algorithm support for the early warning system of economic forest insect pests.

Key words: economic forest, insect pests, YOLOv5, deep learning, feature fusion, object detection