Accurate recognition of jujube tree trunks based on CLAHE image enhancement and improved YOLOv8

Abstract

Background Agricultural image acquisition and target detection are the key links of agricultural precision and intelligence. Facing the practical problems of complex orchard environment and large workload, the existing target detection models have problems such as large number of parameters, slow detection speed, low detection accuracy and poor generalization. Methods In this paper, an improved YOLOv8 target detection model facing the complex environment of orchards is proposed. Firstly, the dataset is efficiently extracted using the key frame extraction algorithm, and the dataset is enhanced using the CLAHE image enhancement method; secondly, the backbone network of the YOLOv8 is replaced with the GhostNetv2 architecture, the GhostConv and GhostBottleneck modules are utilized for the lightweight modification. In addition, the CA_H attention mechanism is improved and added to the Neck part of YOLOv8 network architecture to form YOLOv8s-GhostNetv2-CA_H target detection model. Finally, the effectiveness of the improved model is verified using enhanced dataset comparison tests, ablation tests, and comparison tests with classical methods such as Faster R-CNN and YOLOv5s. Results The average precision of the enhanced dataset over the original dataset rises from 81.2–90.1%; the YOLOv8s-GhostNetv2-CA_H model proposed in this paper reduces the model size by 19.5% compared to the YOLOv8s base model, precision increased by 2.4–92.3%, recall increased by 1.4%, "mAP@0.5" increased by 1.8%, and FPS is 17.1% faster. Conclusions In the complex environment of jujube garden image enhancement has a greater impact on target detection accuracy enhancement, and, the performance of the method proposed in this paper is better than other related algorithms in practical applications.

YOSBG: UAV image data-driven high-throughput field tobacco leaf counting method

“…• The GhostNet [31][32] and C2f modules are merged to form the C2fGhost module [33]. Integrating with the neck network segment of YOLOv8 not only reduces the number of model Params by 16.7% but also improves the overall detection accuracy of the model.…”