by Zhongjian Xie, Yaya Zhang, Weilin Wu, Yao Xiao, Xinwei Chen, Weiqi Chen, ZhuXuan Wan, Chunhua Lin
To achieve automated harvesting of hydroponic Chinese flowering cabbage, the detection and localization of the cabbage are crucial. This study proposes a two stages detection and localization algorithm for hydroponic Chinese flowering cabbage, which includes macro-detection and micro-localization. The macro-detection algorithm is named P-YOLOv5s-GRNF. Its improvement strategies include adopting pruning techniques, the GSConv, receptive field attention convolution (RFAConv), normalization-based attention module (NAM), and the Focal-EIOU Loss module. The micro-localization algorithm is named YOLOv5s-SBC. Its improvement strategies include adding a 160×160 detection layer, removing a 20×20 detection layer, introducing a weighted bidirectional feature pyramid network (BiFPN) structure, and utilizing the coordinate attention (CA) mechanism. The experimental results showed that P-YOLOv5s-GRNF increased the mAP(mean average precision) by 0.8%, 4.3%, 3.2%, 0.7%, 19.3%, 9.8%, 3.1% compared to mainstream object detection algorithms YOLOv5s, YOLOv6s, YOLOv7-tiny, YOLOv8s, YOLOv5s-Shufflenetv2, YOLOv5s-Mobilenetv3, YOLOv5s-Ghost, respectively. Compared to the original model, P-YOLOv5s-GRNF decreased parameters by 18%, decreased model size to 11.9MB, decreased FLOPs to 14.5G, and increased FPS by 4.3. YOLOv5s-SBC also increased mAP by 4.0% compared to the original YOLOv5s, with parameters decreased by 65%, model size decreased by 60%, and FLOPs decreased to 15.3G. Combined with a depth camera, the improved models construct a positioning system that can provide technical support for the automated and intelligent harvesting of Chinese flowering cabbage.