【摘要】 针对光伏板缺陷检测容易产生漏检、误检和检测精度差等问题，为提高缺陷检测的准确性，提出一种基于改进YOLOv11n的光伏板红外图像缺陷检测算法DMDI-YOLOv11n。首先，在主干网络中将C3k2与DRBconv模块融合，提高模型的检测性能的同时减少模型参数量和计算量；其次，在特征融合后并行MLCAttention注意力机制，增强小目标的表达；之后，重构检测头，将多分支、多尺度思想与重参数化思想结合，提高单一卷积的特征提取能力，并降低原有的推理成本；最后，使用Inner＿DIoU损失函数代替CIoU损失函数，弥补边界框回归方法的不足，进一步提高检测能力。实验结果表明：与YOLOv11n模型相比，改进后的算法DMDI-YOLOv11n, mAP50(B)达到87%,与原模型相比提升了1.8%,mAP50-95(B)提升了2.6%;模型的参数量由原来的2.58 M下降至2.17 M,降低了15.89%,准确率提升了4.4%;计算量GFLOPs由6.3 G下降至5.1 G,降低了19.1%,证明改进后的算法DMDI-YOLOv11n能满足光伏板缺陷检测巡检的准确性和边缘部署的轻量化要求。更多还原

【Abstract】 Aiming at the problems such as missed detection, false detection and poor detection accuracy that are prone to occur in the defect detection of photovoltaic panels, in order to improve the accuracy of defect detection, a defect detection algorithm for infrared images of photovoltaic panels based on the improved YOLOv11n, DMDI-YOLOv11n, is proposed. Firstly, C3k2 is fused with the DRBconv module in the backbone network, which improves the model detection performance while reducing the number of model parameters and computation. Secondly, after feature fusion, the MLCAttention attention mechanism is parallelized to enhance the expression of small targets. Subsequently, the detection head is reconstructed to combine the multi-branching and multi-scale ideas with the re-parameterization idea to improve the feature extraction capability of a single convolution and to reduce the original inference cost. Lastly, the Inner＿DIoU loss function is used instead of the CIoU loss function to make up for the deficiencies of the bounding box regression method and further improve the detection ability. The experimental results show that compared with the YOLOv11n model, the improved algorithm DMDI-YOLOv11n, mAP50(B) reaches 87%,representing an increase of 1.8% over the original model, while mAP50-95(B) is improved by 2.6%; the number of parameters of the model decreases from the original 2.58M to 2.17M,representing a reduction of 15.89%; and the accuracy rate is improved by 4.4%. The computational amount GFLOPs decreases from 6.3G to 5.1G,representing a reduction of 19.1%,proving that the improved DMDI-YOLOv11n algorithm can meet both the accuracy requirements for PV panel defect inspection and the lightweight demands for edge deployment.更多还原