Apple inflorescence recognition of phenology stage in complex background based on improved YOLOv7

Accurate discrimination of apple inflorescence morphology and phenology spatial information distribution of orchard are beneficial to guide chemical spraying of target variables and individual thinning operations of machines. In this study, we propose a recognition method based on an improved YOLOv7 model for detecting apple inflorescence at the bud, initial flowering, and full-bloom flowering stages. To reduce parameters, the Efficient Layer Aggregation Network (ELAN) in YOLOv7 was replaced by a residual network structure containing three convolutional layers. A Squeeze and Excitation Network (SENet) and a Coordinate Attention (CA) were embedded in the last layer of the backbone network and the head network, respectively, to improve the recognition accuracy and sensitivity of apple inflorescence. To more accurately compute the distance between the prediction box and the ground truth box. SIoU bounding box regression loss function was used to refine the regression inference bias and improve the bounding box prediction accuracy. In the detection head network, an 80 × 80 detection head was added to improve the recognition ability of small-scale apple inflorescence. Finally, a phenology apple inflorescence dataset was established for the experiment. The ablation experiment results showed that a proper trick could bring an additional performance improvement to the model. The proposed model outperformed three models proposed in the previous study (YOLOv5s, improved YOLOv5, and YOLOv7). It obtained the best performance with a mAP of 80.1% and a recognition speed of 42.58 frames per second (fps). The practicability and robustness of the recognition method were verified by developing the phenology apple inflorescence detection and recognition system. This finding can provide a theoretical basis and strategy for developing real-time recognition equipment for apple inflorescence during phenology.