基于DDR GAN的低质量图像增强算法

Low-quality image enhancement algorithm based on DDR GAN

现有图像增强方法在处理模糊且分辨率较低的图像时,因图像的细节缺乏真实性并且存在伪影现象,会导致增强效果较差。为了解决这一问题,采用一种基于深度密集残差生成对抗网络(DDR GAN)的低分辨率模糊图像增强算法,实现了低质量图像的有效增强。首先构建端到端的生成对抗网络框架;然后设计深度密集残差隐特征编码架构,提升对输入图像的深层语义特征表示,增强图像生成效能;最后重构损失函数,添加感知损失以指导模型学习生成图像的真实性。结果表明,相比于目前最先进的增强型超分辨率GAN法(ESR GAN)和第2版去模糊GAN法(DeBlur GAN-V2),DDR GAN生成的图像在视觉效果上更佳,具有更高的清晰度和更丰富的图像细节;在客观评价指标方面,DDR GAN相较于ESR GAN和DeBlur GAN-V2,峰值信噪比分别提高1.7072 dB和1.1683 dB,结构相似度分别提高0.0783和0.0713。该算法对低分辨率模糊图像的复原增强是有帮助的。

In order to solve the problems of artifacts and unreal details when process low-resolution blurred images using image enhancement methods,a low-resolution blurred image enhancement algorithm based on deep dense residual generative adversarial network(DDR GAN)was used to achieve effective enhancement of low-quality images.Firstly,an end-to-end generative adversarial network framework was constructed;further,a deep dense residual latent feature encoding architecture was designed to improve the deep semantic feature representation of the input image and enhance the image generation efficiency;finally,the loss function was reconstructed and the perceptual loss was added to guide the model to learn to generate realistic images.The comparative experimental results show that compared with the current state-of-the-art enhanced super-resolution GAN(ESR GAN)and DeBlur GAN-V2 algorithms,the images generated by DDR GAN are visually better,with higher definition and richer image detail.In terms of objective evaluation indicators,compared with ESR GAN and DeBlur GAN-V2,the peak signal-to-noise ratio was improved by 1.7072 dB and 1.1683 dB through DDR GAN,respectively,and the structural similarity by 0.0783 and 0.0713,respectively.This algorithm is helpful for the restoration and enhancement of low-resolution blurred images.