Abstract:Apple tree thinning is an important step in orchard production management. Accurate and efficient recognition of apple king flowers and side flowers is the premise of the development of intelligent flower thinning robot. According to the actual demand of apple flower thinning, a method for recognizing king flowers and side flowers of apple based on CRV-YOLO was proposed. Based on YOLO v5s model, the following improvements were made: firstly, C-CoTCSP structure was integrated into Backbone to better learn contextual information and improve the detection performance of the model for king flowers and side flowers that were similar and the position relationship was not obvious. Then an improved RFB structure was added to the Backbone, with which the receptive field of feature extraction was expanded and the branch contribution degree was weighted to make better use of different scale features. Finally, VariFocal Loss loss function was used to improve the detection ability of the model for samples in occlusion and other scenes. Experiments were conducted on a dataset of 1837 images from three varieties. The results showed that the precision, recall and mAP of the proposed model were 95.6%, 92.9% and 96.9%, respectively, which were 3.7 percentage points, 4.3 percentage points and 3.9 percentage points higher than those of the original model. The model was less affected by light changes and apple varieties. Compared with that of Faster R-CNN, SSD, YOLOX, and YOLO v7, precision, the mAP and model size and complexity performance of CRV-YOLO were optimal, and recall was close to optimal. The research results can provide technical support for apple intelligent flower thinning.