Table grape inflorescence detection and clamping point localisation based on channel pruned YOLOV7-TP

Inflorescence thinning is one of the most critical techniques for producing high-quality table grapes. Currently, manual thinning of grape inflorescences is still adopted in most vineyards, and it is hard to reach the demands of large-scale and mechanical planting, which easily misses the best season for thinning inflorescences. It is urgent to design a machine for inflorescence thinning to avoid repetitive labour. The thinning machine re-quires the detection of table grape inflorescences and the location of stem clamping point. A lightweight channel pruned YOLOV7-TP that can be deployed in the consumer-grade vision system of the thinning machine was proposed. First, grape inflorescences and stems with different numbers of clamping points were stacked as labels and fed into the YOLOV7-TP for training. Then an optimal model was obtained after studying the factors impacting the detection and localisation performance. Finally, channel pruning was applied to YOLOV7-TP to compress the model, and some factors affecting the results were analysed for balancing detection accuracy and speed. Experiments were conducted to evaluate the effectiveness of YOLOV7-TP and the optimal combination of hyperparameters was determined. Furthermore, when the sparsity rate was 0.00025 and the pruning rate was 0.4, the channel pruned YOLOV7-TP performed optimally. The model achieves 91.5% mAP0.5, 2.3M parameters, 8.7 GFLOPS, and 29.4 fps detection speed. The channel pruned YOLOV7-TP has demonstrated excellent detection accuracy and speed in detecting grape inflorescence and location stem clamping points simultaneously, which supports a powerful technology for the vision system of grape inflorescence thinning machine. (c) 2023 IAgrE. Published by Elsevier Ltd. All rights reserved.