Resumen
Accurate litchi identification is of great significance for orchard yield estimations. Litchi in natural scenes have large differences in scale and are occluded by leaves, reducing the accuracy of litchi detection models. Adopting traditional horizontal bounding boxes will introduce a large amount of background and overlap with adjacent frames, resulting in a reduced litchi detection accuracy. Therefore, this study innovatively introduces the use of the rotation detection box model to explore its capabilities in scenarios with occlusion and small targets. First, a dataset on litchi rotation detection in natural scenes is constructed. Secondly, three improvement modules based on YOLOv8n are proposed: a transformer module is introduced after the C2f module of the eighth layer of the backbone network, an ECA attention module is added to the neck network to improve the feature extraction of the backbone network, and a 160 × 160 scale detection head is introduced to enhance small target detection. The test results show that, compared to the traditional YOLOv8n model, the proposed model improves the precision rate, the recall rate, and the mAP by 11.7%, 5.4%, and 7.3%, respectively. In addition, four state-of-the-art mainstream detection backbone networks, namely, MobileNetv3-small, MobileNetv3-large, ShuffleNetv2, and GhostNet, are studied for comparison with the performance of the proposed model. The model proposed in this article exhibits a better performance on the litchi dataset, with the precision, recall, and mAP reaching 84.6%, 68.6%, and 79.4%, respectively. This research can provide a reference for litchi yield estimations in complex orchard environments.