Location of safflower filaments picking points in complex environment based on improved Yolov5 algorithm

Mechanized safflower harvesting is prone to inaccurate recognition and positioning of safflower filaments, which is influenced by complex environmental factors such as occlusion, lighting, and challenges related to small targets and small samples. To solve this problem, we improved on the Yolov5 algorithm model and developed a two-stage recognition and positioning approach named Yolov5-ABBM. A safflower dataset was established to classify safflower filaments based on their maturity levels. The Swin Transformer attention mechanism was incorporated to improve the feature-extraction capability of the algorithm model, particularly for small samples and small targets. A geometric operation algorithm based on Bbox and Mask (ABBM) was developed to enhance the positioning speed and minimize missed recognition when locating safflower-filament picking points. Experimental results show that the improved model achieved a recognition precision improvement of 5.8% and 7.9% based on Bbox and Mask, respectively, and exhibited a significant enhancement of 15.3% and 19.4% for small samples. The positioning precision reached 98.19%, with an average positioning running time of 0.018 s per frame image. The improved model demonstrated superior accuracy and positioning speed compared with other algorithm models. The results show that the improved model could accurately identify and locate safflower-filament picking points, particularly for small samples, thereby offering technical support for efficient mechanized safflower harvesting.