Intelligent recognition of citrus shoot growth based on improved YOLOX-Nano algorithm

Objective  In order to solve the problem of low recognition accuracy due to similar color of the background and new shoots, to realize the automatic monitoring of citrus shoot stage and explore the improved method of algorithm, the machine vision technology was used to carry out the research on intelligent perceiving the growth stage of citrus shoot.

Method  According to the characteristics of features extracted from different convolutional layers and the role of different attention mechanism, an improved YOLOX-Nano intelligent recognition model based on multi-attention mechanism was proposed, and a diversified orchard dataset for pre-training was established.

Result  The improved YOLOX-Nano algorithm achieved the mAP (Mean average precision) of 88.07% using the orchard dataset as a pre-training dataset. Compared with the model of YOLOV4-Lite series, the improved model significantly improved the recognition accuracy with less parameters and calculation. Compared with YOLOV4-MobileNetV3 and YOLOV4-GhostNet, the mAP of improved model increased by 6.58% and 6.03% respectively.

Conclusion  The improved model has greater advantage for lightweight deployment at orchard monitoring terminal. The findings provide feasible data and technical solutions for agricultural real-time perception and intelligent monitoring.