Abstract: The automatic detection of fruits is a key technology in agricultural applications such as automatic picking, orchard spraying, and post-harvesting sorting. Aiming at the problems of small citrus targets, many noises, and serious occlusion in the orchard environment, this paper proposes an improved fast identification method for citrus in the orchard environment based on the YOLOv4 algorithm. The main improvement include: one is to use the Canopy algorithm and K-Means++ algorithm to automatically select the number and size of the priori boxes in the training phase; the other is to add an adjustment layer before each output layer of different scale features in the YOLOv4 network, where the residual network structure is combined with densely connected network, and the loss function of the regression box is modified to detect small citrus in a complex background; third, on the premise of ensuring that a large amount of detection accuracy is not lost, the unimportant channels and networks in the network Layers are pruned. The experimental results of comparison with the three commonly used target detection algorithms show that the improved YOLOv4 detection method in this paper has better detection results for citrus in different growth periods in the orchard environment, with an average accuracy rate of 96.04% and a real-time detection speed of 0.06 s per image, are better than the above three mainstream target detection algorithms. The method proposed in this paper can provide technical and methodological guidance for citrus harvesting and yield evaluation in orchards under natural conditions.

Key words: citrus recognition, small target detection, deep learning, improved YOLOv4, convolutional neural networks