Multi-stage Citrus Detection based on Improved Yolov4

Abstract

At present, the research of Citrus recognition is basically aimed at the detection of Citrus in mature stage. This paper proposes a citrus detection algorithm based on improved yolov4, which can detect citrus in each growth stage. Based on yolov4, Introducing CBAM attention mechanism to improve the feature extraction ability of backbone networks; Increase the 22nd layer output of feature extraction network to improve the small target detection rate; A short connection feature fusion method is designed to increase the utilization of shallow feature information; Add a detection head with a scale of 152 * 152 for small-scale targets. It is proved by experiments on the self-built citrus data set, the improved CBAM-F-YOLOv4 can effectively detect citrus in each stage, and the mean Average Precision (mAP) is 6.2 percentage points higher than the original algorithm, reaching 87.3%. The detection results show that the improved algorithm greatly improves the detection ability of occlusion、 overlap and small-scale citrus.