基于改进YOLOv11n的干红辣椒外观品质分级方法OA

The appearance quality of dried hot peppers determines their market value.However,traditional color sorters cannot identify special defects such as broken chili and chili with stems,and the equipment cost is high.Small and medium-sized processing factories urgently need low-cost,edge-deployable intelligent quality grading solutions.To address this,this paper proposed a lightweight method for grading dried hot peppers based on improved YOLOv11n.First,redundant large-object detection heads were removed to construct a dual-head architecture based on the characteristic of concentrated target sizes in chili.Second,the C3k2 modules in the backbone and feature fusion layers were replaced with a self-designed lightweight Faster_C2 module based on partial convolution(PConv),coupled with the adoption of a differentiated channel configuration to reduce computational complexity.Third,dynamic point sampling DySample underwent substitution for static interpolation,enhancing the restoration capability for minute defects such as chili stems and cracks.Finally,the incorporation of the parameter-free SimAM attention mechanism served to suppress background interference from the conveyor belt without increasing the parameter count.Experimental results demonstrated that the improved model achieved an mAP@0.5 of 98.0％,representing a 1.2 percentage point improvement over the original model whose size experienced a reduction to 1.00 MB with the computational complexity decreased to 3.4 GFLOPs,corresponding to reductions of 61.2％and 46.9％,respectively.The improved model was deployed on an Orange Pi 5 Plus embedded board and integrated with a self-developed grading system for validation.The system achieved a stable grading accuracy above 95％for acceptable peppers at a throughput of 36 kg/h,and the CPU temperature maintains below 55 ℃ and the frame rate sustains at 58 FPS during continuous ten hours operation,meeting the requirements for real-time and stable performance in industrial settings.