High-efficiency tea shoot detection method via a compressed deep learning model

Achieving high-efficiency and accurate detection of tea shoots in fields is essential for robotic plucking. A real-time shoot method using the channel layer pruned YOLOv3-SPP deep learning algorithm was proposed this study. First, images were collected data augmentation performed to increase sample diversity, then a spatial pyramid pooling module added YOLOv3 model detect shoots. To simplify improve speed, pruning used compress model. Finally, fine-tuned restore its accuracy, achieve fast The test results demonstrated that number parameters, size, inference time after compression reduced by 96.82%, 96.81%, 59.62%, respectively, whereas mean average precision only 0.40% lower than original In field test, compressed deployed on Jetson Xavier NX conduct experimental speed 15.9 fps, which 3.18 times All indicate could be harvesting robots with low computing power high efficiency detection. Keywords: learning, detection, compression, DOI: 10.25165/j.ijabe.20221503.6896 Citation: Li Y T, He L Y, Jia J M, Chen N, Lyu J, Wu C Y. High-efficiency via Int Agric & Biol Eng, 2022; 15(3): 159–166.