基于归一化控制器的光伏图像无监督域适应缺陷检测

NORMALIZED CONTROLLER FOR PHOTOVOLTAIC DEFECT DETECTION IN DYNAMIC OPEN SCENARIOS

为解决在太阳电池数据集域偏移问题,提出一种针对动态开放场景下光伏缺陷检测的数据归一化控制器(DNC),以提高网络的域适应能力。在测试阶段,该文的DNC方法通过修改模型参数,根据小批次样本数据(0.5%)来实现领域统计纠正。DNC可将目标域产生域偏移的数据映射到与源域数据相同的分布空间,而无需提前标注数据或访问目标域的全部数据。实验结果表明,DNC能显著提高目标检测模型对域偏移数据的适应能力。仅使用少量未标记的目标域数据(少于0.5%)就可在分布外数据上获得显著的性能提升,同时不会降低模型的预测速度(FPS)。

To address domain shift issues in solar cell datasets,this thesis proposes a data normalization controller(DNC)tailored for photovoltaic defect detection in dynamic open environments,aiming to enhance the network’s domain adaptation capability.During the testing phase,the DNC method in this study adjusts model parameters based on small batches of sample data(less than 0.5%),effectively rectifying domain statistics.DNC maps data experiencing domain shift in the target domain onto the same distribution space as the source domain data,without requiring prior labeling of data or access to the entire target domain dataset.Experimental results demonstrate that DNC significantly improves the target detection model’s adaptability to domain-shifted data.Using only a minimal amount of unlabeled target domain data(less than 0.5%),substantial performance gains can be achieved on out-of-distribution data,while maintaining the model’s prediction speed(FPS).