基于双维度注意力集成对抗网络的太阳能电池缺陷图像生成
Solar Cell Defect Image Generation Based on Dual-Dimensional Attention Integrated Adversarial Network
周颖 1裴盛虎 2陈海永 1颜毓泽2
作者信息
- 1. 河北工业大学人工智能与数据科学学院 天津 300130;河北省控制工程技术研究中心 天津 300130
- 2. 河北工业大学人工智能与数据科学学院 天津 300130
- 折叠
摘要
针对太阳能电池缺陷图像稀缺问题,为了对太阳能电池缺陷检测模型进行训练,提出一种双维度注意力集成对抗网络的缺陷图像生成方法.首先构造双生成器与双判别器的集成对抗网络模型;然后将通道注意力与改进的空间注意力结合为双维度注意力,并将其融入生成器与判别器中;最后设计双生成器分时训练的方式解决模型训练不稳定的问题.在太阳能电池电致发光(EL)缺陷数据集上的实验结果表明,5种生成缺陷图像中的图像多样性指标和结构相似性指标比现有最优生成方法最高分别提升53.87和0.46;利用生成的缺陷图像进行yolov5检测模型的训练,5种缺陷的平均精度均值达到96.56％.
Abstract
Aiming at the problem of scarcity of solar cell defect images,a dual-dimensional attention inte-grated adversarial network for defect images generation is proposed for solar cell defect detection model training.Firstly,an integrated adversarial network model with dual generators and dual discriminators is constructed;secondly,channel attention and improved spatial attention are combined into dual-dimensional attention which is incorporated into generators and discriminators;finally,for solving unstable problems when training model,a dual-generator time-sharing training approach is designed.Compared with the exist-ing optimal generation methods on the solar cell electroluminescence(EL)defect dataset,the Fréchet incep-tion distance(FID)and structural similarity index measure(SSIM)of the five kinds generated defect images are increased by 53.87 and 0.46.In addition,the mean average precision(MAP)of the five kinds defect im-ages reaches 96.56％by using the generated defect images to train the yolov5 detection model.
关键词
生成对抗网络/注意力机制/双生成器/双判别器/太阳能电池Key words
generative adversarial networks/attention mechanism/dual generators/dual discriminators/solar cell引用本文复制引用
基金项目
国家自然科学基金(U21A20482)
国家自然科学基金(62073117)
中央引导地方科技发展资金(206Z1701G)
出版年
2024
NETL
NSTL
万方数据