Abstract:In order to address the problems of low detection accuracy across different rice growth stages and easily missed detection under complex field environments, an improved rotated object detection model AHF-YOLO 11 was proposed based on YOLO 11. This model introduced convolutional vision module AssemFormer into YOLO 11 backbone network to construct C3k2_AssemFormer feature extraction module, significantly enhancing the local feature expression and cross reproductive global morphology learning ability in dense occlusion scenes. Furthermore, it adopted the high-level screening feature pyramid network (HSFPN) to replace the original neck structure, effectively improving the ability of the model to recognize multi-scale rice panicles under varying backgrounds. Experimental results demonstrated that the accuracy of the improved AHF-YOLO 11 model reached 90.7%, which represented an improvement of 4.3 percentage points higher than that of the original YOLO 11 model and 4.4~39 percentage points higher than the mainstream models, respectively. The results of cross-growth-stage testing further revealed that the recognition accuracy of AHF-YOLO 11 in the booting, heading, filling, and maturity stages was respectively improved by 11.88, 7.3, 4.78, and 4.3 percentage points when compared with that of the original model. Among these, the highest accuracy of 94.41% was achieved for the model at the heading stage. Further follow-up tracking studies and yield estimation experiments indicated that the period from the late heading stage to the early filling stage was the optimal period for rice yield estimation, with the lowest yield estimation error of only 4.66%. The research result can provide important technical support for rice panicle number recognition and yield estimation.
