【摘要】 太阳能电池板晶体表面易碎，在生产和使用中容易出现各种缺陷，及时准确地检测这些缺陷，对于提高光伏组件的使用效率至关重要。传统的检测方法训练一个太阳能电池缺陷检测模型往往需要对大量数据集进行专业标注，需要耗费大量的人力成本；此外，工业环境情况复杂又会导致获取的图像包含较多背景噪声，且不同的应用环境中采集的图像有较大的域间差异。为了减小标注成本，应对域间差异检测困难，采用主动学习的策略来挑选更有价值的样本进行标注，并引入小样本目标检测的元学习思想对模型进行重加权，从而提高模型应对不同目标域的泛化能力。实验结果表明，在夜间域和曝光域数据集上使用少量的样本可以得到比传统方法更好的检测效果。更多还原

【Abstract】 The crystal surface of industrial solar panels is fragile, and various defects are easy to occur in production and use to affect the use of photovoltaic modules, so it is to detect these defects in time.Training a detection model often requires many corresponding data sets for professional annotation, and external factors(such as illumination, visibility, gray layer) in the actual situation of industrial detection will cause some captured images to have large background noise, affecting the detection effect, so the detection model needs to have strong generalization ability.This paper introduces the idea of active learning to select more valuable samples for labeling, and combines the meta-learning idea of few-shot object detection to use these new samples to reweight the model and improve the generalization ability of the model.The experimental results show that using a small number of samples on the night-domain and exposure-domain datasets yields better detection results than traditional methods.更多还原