【摘要】 针对低分辨率电力设备热成像图像，提出一种基于改进边缘注意力生成对抗网络的超分辨率重建方法。首先，在边缘注意力的基础上，引入通道注意力和位置注意力的双注意力模块(dual attention, DA)，捕获特征图不同位置间和不同通道间的依赖关系，并将两组依赖关系进行融合，以加大全局信息的提取程度。然后针对参数修正线性单元激活函数(parametric rectified linear unit, PReLU)对网络中神经元进行无差别激活，导致网络特征表达能力受限问题。采用改进β-ACONC自适应控制激活函数替代PRe LU函数，在辨识有效特征的基础上，对神经元进行选择性激活，以强化有效特征、弱化无效特征，提升网络的自适应激活能力和特征表达能力。最后对所提改进边缘注意力生成对抗网络模型(edge-attention generative adversarial network, EA-GAN)进行实验验证。结果表明，与Bi Cubic双三次插值模型和原EA-GAN模型边缘注意力生成对抗网络模型相比，所提改进模型网络性能最好，重建图像质量最高，客观评价指标峰值信噪比(peaksignal-to-noiseratio, PSNR)均值、结构相似性(structural similarity,SSIM)均值和均方误差损失(mean square error loss, MSE-loss)均值最优，在电力设备红外图像重建领域普适性较高，具有一定的工程应用价值。更多还原

【Abstract】 A super-resolution reconstruction method based on improved edge-attention generation adversarial network is proposed for low-resolution thermal imaging images of power equipment. First, using edge attention, a dual attention(DA) module of channel and position attention is introduced to capture the dependencies between different positions of the feature map and between different channels. The two sets of dependencies are fused to increase the degree of global information extraction. Then, to address the problem that the parametric rectified linear unit(PReLU) activation function performs undifferentiated activation on the neurons in the network, which leads to the limited feature expression capability of the network. The improved β-ACONC function is used to replace the PReLU function and selectively activate the neurons on the basis of identifying the effective features in order to strengthen effective features and weaken the ineffective features, and enhance the adaptive activation and feature expression capabilities of the network. Finally, the proposed improved edge-attention generative adversarial network(EA-GAN) model is experimentally validated. The results show that compared with BiCubic and the original EA-GAN model, the proposed improved model has the best network performance, the highest reconstructed image quality, and the best objective evaluation indices of peak signal-tonoise ratio(PSNR), structural similarity(SSIM) and mean square error loss(MSE-loss) mean values. These are universal in the field of infrared image reconstruction of power equipment and have a certain engineering application value.更多还原