您的位置：首页 > 论文页面
基于深度学习的图像修复算法的研究
发表时间：2023-06-30 浏览量：1777 下载量：106
| 全部作者： | 张宇，张永军 |
| 作者单位： | 北京邮电大学电子工程学院 |
| 摘 要： | 传统的生成对抗网络（generative adversarial network，GAN）对于大面积缺损区域的图像修复效果不佳。为解决这一问题，本文在GAN模型中引入协同调制机制来综合条件GAN能够控制图像生成风格的优势和非条件GAN随机生成图像能力强的优势，使得网络能够对大面积的缺损图像进行有效地修复。同时，传统的逻辑损失函数导致鉴别器的特征表示泛化不足，不能刺激生成器的对抗性进化，容易忘记之前的数据模式和任务，针对这一问题，将采用双重对比损失机制，避免生成样本语义结构受损或生成分布模式失效等问题。通过结合这些方法，提高了峰值信噪比（peak signal-to-noise ratio，PSNR）、结构相似度（structural similarity，SSIM）和FID（Fréchet Inception Distance）指标在图像恢复中的性能。为了提高修复效率，本文使用yolo对缺损图像区域进行检测，并加入盒注意力机制来提高检测精度。 |
| 关 键 词： | 人工智能；生成对抗网络（GAN）；yolo；图像修复 |
| Title： | Research of image inpainting algorithm based on deep learning |
| Author： | ZHANG Yu, ZHANG Yongjun |
| Organization： | School of Electronic Engineering, Beijing University of Posts and Telecommunications |
| Abstract： | The traditional generative adversarial networks (GAN) are not effective in image restoration of large areas of defect. In order to solve this problem, this paper introduces a co-modulation mechanism into the GAN model to synthesize the advantages of conditional GAN’s ability to control the style of image generation and the advantages of strong ability of non-conditional GAN to randomly generate images, so that the network can effectively repair large-area defective images. At the same time, the traditional logical loss function leads to insufficient generalization of the feature representation of the discriminator, which cannot stimulate the adversarial evolution of the generator, and it is easy to forget the previous data mode and task. To solve this problem, the double contrast loss mechanism will be used to avoid the damage of the semantic structure of the generated sample or the failure of the generated distribution mode. By combining these methods, the performance of peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and Fréchet Inception Distance (FID) indicators in image recovery is improved. In order to improve the repair efficiency, this paper uses yolo to detect the defective image area and adds a box attention mechanism to improve the detection accuracy. |
| Key words： | artificial intelligence; generative adversarial network (GAN); yolo; image inpainting |
| 发表期数： | 2023年6月第2期 |
| 引用格式： | 张宇，张永军. 基于深度学习的图像修复算法的研究[J]. 中国科技论文在线精品论文，2023，16（2）：160-169. |
4
评论数 0请您登录
暂无评论
下载全文
多维论文