Scholarlink

<p>Rainy or hazy weather conditions result in significant quality loss and image blurring for captured images. Restoration through Deep Learning (DL) techniques which include CNNs and GANs proves successful but high-resolution images present a challenge for existing systems because they require excessive computation and their models cannot handle complete image features. The paper introduces a new solution through a hybrid architecture which combines Self-Attention GAN (SAGAN) and weighted Deep CNN technology. The SAGAN component enables long-range dependency capture across the entire image, while a Greedy-based Genetic Algorithm optimizes feature extraction to reduce redundancy. The U-Net architecture serves dual functions of spatial segmentation and feature representation enhancement while it preserves structural integrity. The model was assessed through two datasets which included the RESIDE dataset for dehazing and Rain 100H for deraining. The experimental outcomes demonstrate promising performance through a Peak Signal-to-Noise Ratio (PSNR) measurement of 31.23 dB and a Structural Similarity Index (SSIM) score of 0.999. The proposed system achieves high computational efficiency while operating with 0.155&#xa0;M parameters and 120.75 GFlops, which enables it to surpass existing methods both in restoration quality and resource management. This framework provides a better solution for operational purposes in real-time digital photography and medical imaging environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Generative adversarial network based deraining and Deep CNN for Image Dehazing

  • 期刊论文

  • 发表日期：

    2026年6月25日

查看全文
  • Poornima Shrivastava,
  • Roopam Gupta,
  • Asmita A. Moghe

摘要

Rainy or hazy weather conditions result in significant quality loss and image blurring for captured images. Restoration through Deep Learning (DL) techniques which include CNNs and GANs proves successful but high-resolution images present a challenge for existing systems because they require excessive computation and their models cannot handle complete image features. The paper introduces a new solution through a hybrid architecture which combines Self-Attention GAN (SAGAN) and weighted Deep CNN technology. The SAGAN component enables long-range dependency capture across the entire image, while a Greedy-based Genetic Algorithm optimizes feature extraction to reduce redundancy. The U-Net architecture serves dual functions of spatial segmentation and feature representation enhancement while it preserves structural integrity. The model was assessed through two datasets which included the RESIDE dataset for dehazing and Rain 100H for deraining. The experimental outcomes demonstrate promising performance through a Peak Signal-to-Noise Ratio (PSNR) measurement of 31.23 dB and a Structural Similarity Index (SSIM) score of 0.999. The proposed system achieves high computational efficiency while operating with 0.155 M parameters and 120.75 GFlops, which enables it to surpass existing methods both in restoration quality and resource management. This framework provides a better solution for operational purposes in real-time digital photography and medical imaging environments.