Defect detection for photovoltaic (PV) cell images is a challenging task due to the small size of the defect features and the complexity of the background characteristics. Modern detectors rely mostly on proxy learning objectives for prediction and on manual post-processing components. One-to-one set matching is a critical design for DEtection TRansformer (DETR) in order to provide end-to-end capability, so that does not need a hand-crafted Efficient Non-Maximum Suppression NMS. In order to detect PV cell defects faster and better, a technology called the PV cell Defects DEtection Transformer (PD-DETR) is proposed. To address the issue of slow convergence caused by DETR’s direct translation of image feature mapping into target detection results, we created a hybrid feature module. To achieve a balance between performance and computation, the image features are passed through a scoring network and dilated convolution, respectively, to obtain the foreground fine feature and contour high-frequency feature. The two features are then adaptively intercepted and fused. The capacity of the model to detect small-scale defects under complex background conditions is improved by the addition of high-frequency information. Furthermore, too few positive queries will be assigned to the defect target via one-to-one set matching, which will result in sparse supervision of the encoder and impair the decoder’s ability of attention learning. Consequently, we enhanced the detection effect by combining the original DETR with the one-to-many matching branch. Specifically, two Faster RCNN detection heads were added during training. To maintain the end-to-end benefits of DETR, inference is still performed using the original one-to-one set matching. Our model implements 64.7% AP on the PVEL-AD dataset.

由于缺陷特征尺寸小且背景特征复杂，光伏电池图像的缺陷检测是一项具有挑战性的任务。现代检测器主要依赖代理学习目标进行预测和手动后处理组件。一对一的集合匹配是 DEtection TRansformer (DETR) 的关键设计，以提供端到端功能，因此不需要手工制作的高效非极大值抑制 NMS。为了更快更好地检测光伏电池缺陷，提出了一种称为光伏电池缺陷检测变压器（PD-DETR）的技术。为了解决 DETR 将图像特征映射直接转换为目标检测结果导致的收敛速度慢的问题，我们创建了一个混合特征模块。为了实现性能和计算之间的平衡，图像特征分别经过评分网络和扩张卷积，以获得前景精细特征和轮廓高频特征。然后，这两个特征被自适应地拦截和融合。通过高频信息的加入，提高了模型在复杂背景条件下检测小尺度缺陷的能力。此外，通过一对一的集合匹配将很少的正查询分配给缺陷目标，这将导致编码器的稀疏监督并削弱解码器的注意力学习能力。因此，我们通过将原始DETR与一对多匹配分支相结合来增强检测效果。具体来说，在训练时添加了两个 Faster RCNN 检测头。为了保持DETR的端到端优势，仍然使用原始的一对一集合匹配来进行推理。我们的模型实现了 64。PVEL-AD 数据集上的 AP 为 7%。