基于Tw_Cycle Gan的绝缘子缺陷样本自动生成技术

Automatic generation technology of insulator defect samples based on Tw_Cycle Gan

生成对抗网络(Gan)被应用于电力巡检缺陷样本生成工作以解决缺陷样本不足问题。目前基于Gan的绝缘子缺陷样本生成技术存在如下不足:依赖大量缺陷样本训练且生成量不足;生成质量较差,尺寸较小,难以供目标检测神经网络训练使用。针对上述问题,提出一种基于Starganv2的风格迁移目标加权循环一致(Tw_Cycle)Gan网络,其可借助非缺陷样本训练,并依据非缺陷样本实现一对多缺陷样本生成。为保证缺陷语义不变,加入Unet分割网络,使用目标循环一致及目标掩码损失加强绝缘子目标物的约束。通过定性与定量评估,Tw_Cycle Gan取得了更好的结果。为了验证生成样本的有效性,设计了一种基于真实样本的缺陷检测实验评估方法。结果表明,使用生成缺陷样本扩增训练的同一YOLOv3目标检测算法,AP平均提升5%左右,Precision平均提升4.6%左右,Recall平均提升10%左右,F1平均提升0.083。

Generative adversarial network(Gan) is applied to the generation of defect samples for power inspection to solve the problem of insufficient defect samples. The current Gan-based technology of insulator defect sample generation has the following limitations: a large number of defect samples are required for training and the number of generations is insufficient;the quality of the generated samples is poor, the size is small, making it difficult to use for target detection neural network model training. To address the above limitations, a style transfer a target weighted Cycle consistent(Tw_Cycle) Gan is proposed based on Starganv2. The network can use non-defective samples for training, and realize one-to-many defect sample generation based on non-defective samples. In order to ensure the semantics of defects remain unchanged, the Unet segmentation network is added, and the Roi_cyc Loss and the Roi_mask Loss are used to strengthen the constraint of the insulator target. Through qualitative and quantitative evaluation, Tw_Cycle Gan has achieved better results. In order to verify the validity of the generated samples, an experimental evaluation method for defect detection based on real samples is designed. The results show that the same YOLOv3 target detection algorithm that uses synthetic defect samples to amplify training, the AP, increased by about 5% on average, Precision increased by about 4.6% on average, Recall increased by about 10% on average, and F1 increased on average by 0.083.