Spikelets detection of table grape before thinning based on improved YOLOV5s and Kmeans under the complex environment

Inflorescence thinning is the primary method for crop regulation to obtain high-quality table grapes in viticulture. It is essential to reduce labor dependency and associated costs by using mechanical thinning. To achieve precision mechanical thinning, visual detection of table grape inflorescence and spikelet is an important part of the process. In this article, an end-to-end method based on the improved YOLOV5s and Kmeans algorithm under the complex growing environment is proposed to detect spikelets needed to be removed from the table grape inflorescence. Firstly, the following improvements are made in the YOLOV5s: (1) The attention mechanism Pyramid Split Attention (PSA) establishes longer-distance channel dependencies. (2) The Bi-directional Feature Pyramid Network (BiFPN) enhances the multi-scale feature fusion. (3) CIoU loss makes more accurate regression of bounding box. Then, the test set is input into the improved YOLOV5s model to obtain the predicted bounding box of inflorescences and spikelets. The inflorescences are matched to the spikelets on them by the IoU function, and the Kmeans algorithm is used to cluster the center coordinates of the matched spikelet bounding boxes and determine the tail of the inflorescence according to the aggregation degree. 2/3 of the spikelet bounding boxes on the tail of inflorescence are taken as the spikelets removal. Finally, experiments are designed to verify the detection performance of the proposed method. What’s more, compared with the original YOLOV5s, the improved model has 4.4 percentage points higher mAP value and 7ms slower detection speed than the original YOLOV5s, but still within the acceptable range. Compared with the Faster R-CNN, Cascade R-CNN, SSD, Retinanet, YOLOV3, and YOLOX-s, the improved YOLOV5s improves by 8, 7, 3.2, 9.6, 3.4, 1.9 percentage points in mAP, respectively. It indicates that the improved YOLOV5s detection accuracy and speed can reach a high level. In addition, the parameters of the proposed method are experimentally analyzed to determine the optimal parameters in this article. The results show that the algorithm has better accuracy when the number of spikelets threshold is 20, the IoU threshold is 0.15, the confidence threshold is 0.6, and its maximum accuracy is 78%. Therefore, the proposed algorithm has better detection accuracy and speed under the complex environment and provides theoretical support for the development of table grape thinning machinery.