In traditional photography's limitations, where focus is fixed now of capture, this research uses CycleGAN to enable post-capture focus adjustments, enhancing both the correction of focus errors and the creative modification of images. This study explores the innovative use of Cycle-Consistent Generative Adversarial Networks (CycleGAN) for generating arbitrary focus images, specifically concentrating on leaf imagery. The research employs a two-part neural network architecture inherent in Generative Adversarial Networks (GANs), comprising a generator and a discriminator. These networks are trained concurrently, where the generator aims to produce increasingly realistic images, and the discriminator evaluates their authenticity. CycleGAN’s unique feature of cycle consistency is critical for image-to-image translation tasks, particularly useful in scenarios where paired examples for training are not available, such as in focus manipulation. In this study, CycleGAN is utilized to transform the focus of leaf images by generating dense focal stacks from sparse ones, demonstrating the model’s ability to learn domain transformations from near-focus to far-focus images and vice versa. The process involves subjective evaluations using the naked eye and objective evaluations using metrics like PSNR (Peak Signal-to-Noise Ratio). However, challenges such as significant color tone changes and deterioration in image quality were observed, indicating the model's tendency to learn using color features rather than focus. To enhance image quality, the study incorporates super-resolution methods like VDSR and EDSR, which use residual learning to deepen the model for more efficient training and to mitigate the risk of vanishing gradients. The results demonstrate a substantial improvement in CycleGAN’s accuracy, reaching 97.8%, with PSNR values indicating the model's effectiveness in generating high-quality images.

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

Arbitrary Leaf Image Focus Generation and Enhancement with CycleGAN and Super Resolution

  • V. Grishma Neha Chowdary,
  • Bitta Hari Charan,
  • K. Sree Suryakanth,
  • Raj Kumar Chanda,
  • Pavan Kumar Pagadala,
  • P. Lalitha Surya Kumari

摘要

In traditional photography's limitations, where focus is fixed now of capture, this research uses CycleGAN to enable post-capture focus adjustments, enhancing both the correction of focus errors and the creative modification of images. This study explores the innovative use of Cycle-Consistent Generative Adversarial Networks (CycleGAN) for generating arbitrary focus images, specifically concentrating on leaf imagery. The research employs a two-part neural network architecture inherent in Generative Adversarial Networks (GANs), comprising a generator and a discriminator. These networks are trained concurrently, where the generator aims to produce increasingly realistic images, and the discriminator evaluates their authenticity. CycleGAN’s unique feature of cycle consistency is critical for image-to-image translation tasks, particularly useful in scenarios where paired examples for training are not available, such as in focus manipulation. In this study, CycleGAN is utilized to transform the focus of leaf images by generating dense focal stacks from sparse ones, demonstrating the model’s ability to learn domain transformations from near-focus to far-focus images and vice versa. The process involves subjective evaluations using the naked eye and objective evaluations using metrics like PSNR (Peak Signal-to-Noise Ratio). However, challenges such as significant color tone changes and deterioration in image quality were observed, indicating the model's tendency to learn using color features rather than focus. To enhance image quality, the study incorporates super-resolution methods like VDSR and EDSR, which use residual learning to deepen the model for more efficient training and to mitigate the risk of vanishing gradients. The results demonstrate a substantial improvement in CycleGAN’s accuracy, reaching 97.8%, with PSNR values indicating the model's effectiveness in generating high-quality images.