Real-time monitoring of pepper harvesting loss using a lightweight vision model

● A robust lightweight detection network (Pepper-YOLO) is proposed for visibility-degraded agricultural environments. ● Prior-guided restoration strategy effectively mitigates domain shift caused by heavy dust. ● Multi-scale feature fusion and SimAM attention enhance representation of small, occluded targets. ● Self-Calibrated Illumination Network (SCINet) adapts to complex unstructured lighting conditions. ● Deployed on RDK X5 via PTQ quantization, achieving 18.2 ms real-time inference at the edge. Real-time monitoring of pepper losses during mechanized harvesting is essential for optimizing operational parameters and reducing yield loss. However, accurate detection of fallen peppers in field environments remains challenging due to heavy dust interference and complex backgrounds. To address these issues, this study proposes a robust vision-based monitoring method that integrates image enhancement and lightweight object detection. First, the Dark Channel Prior (DCP) algorithm is applied to reduce dust effects and restore image clarity under degraded field conditions. Then, an improved YOLOv8n-based model, termed Pepper_YOLOv8n, is developed by optimizing the neck structure and incorporating the SimAM attention mechanism together with the Self-Calibrated Illumination Network (SCINet) module, thereby enhancing feature representation for small and partially occluded targets. Experimental results show that the proposed model achieves a precision of 90.1%, recall of 89.7%, and an mAP of 96.8%. Compared with the baseline YOLOv8n, recall and [email protected] were improved by 4.0 and 2.3 percentage points, respectively, while the inference time was reduced by 1.5 ms on a workstation. To further validate its practical applicability, a real-time edge detection system is developed based on the D-Robotics RDK X5 platform. With post-training quantization (PTQ), the deployed model reduces computational complexity (FLOPs decreased by 3.6G) and achieves an inference speed of 18.2 ms per frame, meeting the requirements for real-time monitoring during field harvesting operations. The proposed method provides an effective solution for intelligent monitoring of harvesting losses and offers practical support for precision agriculture and smart agricultural machinery.