摘要
针对太阳能电池片缺陷数据量匮乏造成的网络过拟合和模型性能不达标的问题,提出基于深度卷积对抗生成网络和图像随机拼接的真假数据融合算法,将训练数据量提升了800倍;同时对网络模型进行轻量化优化,减少模型训练参数。实验结果表明,经过真假数据融合扩充数据集后训练的模型测试精度相比原始训练集和传统数据增强算法分别提升了近30%和17%;轻量化处理后的模型参数减少为之前的1/2,对每张图片的测试时间由57 ms缩短到22 ms。研究证明,真假数据融合算法能够有效的缓解训练数据不足造成网络过拟合问题;轻量化优化模型在保证精度的同时,压缩模型大小,加快测试速度。
Aiming at the problem of network overfitting and model performance under standard caused by the lack of defective data amount of solar cells, In this paper, a true and false data fusion algorithm based on deep convolution confrontation generation network and random image Mosaic is proposed, which improves the training data volume by 800 times. At the same time, the network model is optimized with light weight to reduce model training parameters. The experimental results show that the test accuracy of the trained model after the data fusion and expansion of the data set is nearly 30% and 17% higher than that of the original training set and the traditional data enhancement algorithm. After the lightweight treatment, the model parameters were reduced to about half of the previous ones, and the test time for each image was shortened from 57 ms to 22 ms. The research shows that the fusion algorithm can effectively alleviate the problem of network overfitting caused by insufficient training data. The lightweight optimization model not only ensures the accuracy, but also compresses the size of the model to speed up the test.
作者
王云艳
周志刚
罗帅
Wang Yunyan;Zhou Zhigang;Luo Shuai(School of Electrical and Electronic Engineering,Hubei University of Technology,Wuhan 430068,China)
出处
《电子测量与仪器学报》
CSCD
北大核心
2021年第1期26-32,共7页
Journal of Electronic Measurement and Instrumentation
基金
国家自然科学基金(41601394)
湖北工业大学博士启动基金(BSQD2016010)资助项目。
