基于改进RT-DETR的光伏红外缺陷检测算法OA
Photovoltaic infrared defect detection algorithm based on improved RT-DETR
光伏板表面缺陷检测是保障光伏系统发电效率与运行安全的关键.针对无人机航拍红外图像中航拍小目标检测难、户外热背景干扰强及缺陷目标尺度跨度大等挑战,本研究提出一种结合多尺度特征与动态卷积的光伏红外缺陷检测模型.首先,在浅层路径采用空间到深度卷积(SPD-Conv)替代传统下采样,规避二极管异常、热斑等微小目标的细粒度特征与边缘信息丢失;其次,引入动态全维卷积(CSPOmniKernel)模块,通过多分支机制自适应调整感受野,增强低对比度缺陷特征并抑制复杂热背景干扰;最后,构建高效多尺度特征融合模块(MFM),借助跨层级信息交互与自适应聚合,提升模型对不同尺度目标的感知鲁棒性.在包含5类典型缺陷的光伏红外航拍数据集上,改进模型的mAP@0.5达到84.12%,较基准RT-DETR-r18提升5.16%,且推理速度达48 FPS,满足工业实时巡检需求.结果表明,该方案在复杂户外场景下能有效提升光伏缺陷检测的精度与可靠性.
Surface defect detection of photovoltaic panels is of paramount importance for ensuring the power generation efficiency and operational safety of photovoltaic systems.To tackle challenges such as the difficulty in detecting small aerial targets in infrared images captured by drones,significant interference from outdoor thermal backgrounds,and the broad scale range of defect targets,this study introduces a photovoltaic infrared defect detection model that integrates multi-scale features with dynamic convolution.First,in the shallow pathway,spatial-to-depthwise convolution(SPD-Conv)is employed to substitute traditional downsampling methods,thereby preventing the loss of fine-grained features and edge information pertaining to small targets,including diode anomalies and hot spots.Second,a dynamic full-dimensional convolution(CSPOmniKernel)module is incorporated,which adaptively modulates the receptive field through a multi-branch mechanism,thereby enhancing low-contrast defect features while mitigating interference from complex thermal backgrounds.Finally,an efficient multi-scale feature fusion module(MFM)is devised,leveraging cross-level information interaction and adaptive aggregation to bolster the model's robustness in perceiving targets of varying scales.On a photovoltaic infrared aerial dataset encompassing five typical defect types,the enhanced model attains an mAP@0.5 of 84.12%,marking a 5.16%improvement over the baseline RT-DETR-r18,and achieves an inference speed of 48 FPS,fulfilling the demands of industrial real-time inspection.The results demonstrate that this approach significantly elevates the accuracy and reliability of photovoltaic defect detection in intricate outdoor environments.
邓万宇;彭鑫亮
西安邮电大学 计算机学院,陕西 西安 710121西安邮电大学 计算机学院,陕西 西安 710121
信息技术与安全科学
光伏缺陷检测航拍红外图像RT-DETR算法动态卷积小目标检测
photovoltaic defect detectionaerial infrared imageRT-DETR algorithmdynamic convolutionsmall target detection
《液晶与显示》 2026 (8)
1175-1186,12
陕西省科学家工程师创新团队项目(No.2023KXJ-091)秦创原总窗口"四链融合"计划重点项目(No.2025PT-ZCK-67)Supported by Shaanxi Scientist&Engineer Innovation Team(No.2023KXJ-091)Key Project of Qinchuangyuan General Window"Four-Chain Integration"Program(No.2025PT-ZCK-67)
评论