基于循环互相关系数的CGAN温度值图像扩增

针对变电设备红外图像温度值样本少、不均衡等问题，本文提出了一种基于循环互相关系数的条件生成对抗网络(conditional generative adversarial network, CGAN)温度值图像扩增方法。根据图像相似度提出了循环互相关系数，改进了CGAN模型的损失函数；使用改进后的CGAN模型在原始温度值图像数据集的基础上进行图像扩增，得到了包含11种标签的新数据集；采用卷积神经网络对传统图像扩增方法、原始CGAN模型和改进的CGAN模型扩增的图像进行对比和测试。结果表明，改进的CGAN模型收敛速度更快，训练过程稳定，扩增的图像轮廓清晰、细节丰富，客观评价指标最高，温度值识别准确率达到99.4%，实现了图像扩增的目的。

Image amplification for temperature value image based on cyclic cross-correlation coefficient CGAN

To solve the problems of small sample size and imbalance of infrared image for substation equipment, a temperature image amplification method based on cyclic cross-correlation coefficient conditional generative adversarial network (CGAN) is proposed. The cyclic cross-correlation coefficient is proposed according to the image similarity, which improves the loss function of CGAN. Then the improved CGAN is used to amplify the original temperature image data set, establishing a new data set containing 11 labels. Then, the traditional image amplification method, the original CGAN and the improved CGAN are compared using the convolution neural network (CNN). The experiment demonstrate that the proposed CGAN model has faster convergence speed and stable training process, and the generated images have clear contour and rich details. The objective evaluation index of the proposed method is the largest, and the recognition accuracy of temperature value reaches 99.4%, which realizes the purpose of the image amplification.