A lightweight model based on you only look once for pomegranate before fruit thinning in complex environment

Abstract

Using picking robot to thin pomegranate, the accuracy and speed for the algorithm are very significant, especially in complex environments. Therefore, a detection method TP-YOLO (Thinning pomegranate-YOLO) is proposed through model lightweighting and improvement in recognition accuracy based on You Only Look Once Version 8 (YOLOv8s). The lightweighting of the model aspect ShuffleNetV2 is firstly introduced to reconstruct the backbone of YOLOv8s, and the standard convolution of Neck is replaced by depthwise separable convolution. Then the feature level of the model is modified. The improvement in recognition accuracy is mainly achieved by replacing the residual structure of ShuffleNetV2 with ShuffleNetV2-SE, which includes Squeeze-and-Excitatio (SE) attention mechanism. Then, the proposed algorithm is trained and tested with self-built pomegranate dataset before fruit thinning. Moreover, TP-YOLO is embedded into the self-built pomegranate growth status detection platform. The experimental results indicate that the Mean Average Precision (mAP), Size, Giga Floating-point Operations Per Second (GFlops) of TP-YOLO model are 94.4%, 1.9 MB, 8.5, respectively. Furthermore, compared with the latest research results, the number of parameters of our algorithm is reduced by 67.9% while there is no decrease in the detection accuracy. This provides a research foundation for fruit picking robots application to the automation and intelligent development of the pomegranate industry.