【摘要】 在金属凝固的微观组织图像中广泛存在着随机噪声和语义噪声,这些噪声严重干扰了对微观组织的特征提取与定量分析。针对以上问题,文中提出了一个基于生成对抗网络的图像盲增强算法,构建了双阶段生成网络,其中一阶段残差学习网络针对处理随机噪声,其提取的噪声特征可以有效地融合入二阶段图像修复网络中。此外,在修复网络中引入凝固组织的空间结构一致性约束,以提升边缘的重建精细程度。实验表明,文中算法的PSNR与SSIM分别达到39.16dB和0.9937,优于其他典型方法,可用于后续凝固组织的定量分析。更多还原

【Abstract】 Random noise and semantic noise are widely exist in the microstructure image during metal solidification. These noises have seriously affected feature extraction and quantitative analysis of microstructures. To solve the above problem, this paper proposes a blind enhancement algorithm for material images based on generative adversarial networks. A two-stage generative network is built, where the residual learning network at the first stage mainly aims at reducing random noise, and the extracted features of random noise can be effectively integrated into the second stage network. In addition, the consistency constraint of spatial contours from solidification is introduced in the inpainting network to improve the reconstruction accuracy of the edges. The results show that the restored images achieves PSNR value of 39.16 dB and SSIM value of 0.9937, respectively, which is superior to other typical image enhancement methods. The proposed method has broad prospects in quantitative analysis of microstructures.更多还原