Scholarlink

<p>Image denoising plays very crucial role in applications which include photography/videography, Security-Surveillance, Medical Image analysis, Astronomy, Aerial photography, Satellite Imagery, and many more. The results of denoising also affect the accuracy of later tasks of image segmentation, object recognition, classification. Deep learning-based approaches have started addressing this issue for many years. Major challenges with these approaches are the unavailability of ground truth clean images, sensitivity of the model to type/amount of noise over which it is trained, and artifacts produced. Self-supervised approaches play an important role in addressing the first issue. These self-supervised approaches like Noise2Noise, Neighbor2Neighbor, Recorrupted2Recorrupted etc., derive the target, which acts as the Pseudo-clean ground truth. Thus, the underlying base model can be trained by using the noisy and derived clean target images, which does not need the clean ground-truth. The Proposed approach includes bicubic interpolation based self-supervised proposed approach for Image denoising. It takes the noisy image as input and downscales it at multiple levels. After that those downscaled images are again upscaled to the original dimension with the help of bicubic-interpolation. These upscaled images contain lesser noise and more information compared to original noisy one and can be treated as the target Pseudo-clean images. This way, the underlying encoder-decoder based convolutional neural network, U-Net, learns to reconstruct the underlying clean image from its noisy counterpart. Loss function is proposed to include MSE (Mean Squared Error), PSNR (Peak Signal to Noise Ratio), and SSIM (Structural Similarity Measurement) terms to balance between denoising results and preservation of structural information. 5000 Images from the ILSVRC2012 ImageNet validation dataset have been selected for the training of the proposed approach. BSD300, Kodak, Set14, and SIDD datasets have been used for the purpose of evaluation. Experimental results over the datasets show that on an average the proposed method has comparable performance than that of Noise2Noise and Neighbor2Neighbor for the Gaussian, Poisson image denoising scenarios.</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

Self-Supervised Image Denoiser Design Using Multiscale Bicubic Image Interpolation and U-Net Network

  • 期刊论文

  • 发表日期：

    2025年2月18日

查看全文
  • Ashishkumar Gor,
  • C. K. Bhensdadia

摘要

Image denoising plays very crucial role in applications which include photography/videography, Security-Surveillance, Medical Image analysis, Astronomy, Aerial photography, Satellite Imagery, and many more. The results of denoising also affect the accuracy of later tasks of image segmentation, object recognition, classification. Deep learning-based approaches have started addressing this issue for many years. Major challenges with these approaches are the unavailability of ground truth clean images, sensitivity of the model to type/amount of noise over which it is trained, and artifacts produced. Self-supervised approaches play an important role in addressing the first issue. These self-supervised approaches like Noise2Noise, Neighbor2Neighbor, Recorrupted2Recorrupted etc., derive the target, which acts as the Pseudo-clean ground truth. Thus, the underlying base model can be trained by using the noisy and derived clean target images, which does not need the clean ground-truth. The Proposed approach includes bicubic interpolation based self-supervised proposed approach for Image denoising. It takes the noisy image as input and downscales it at multiple levels. After that those downscaled images are again upscaled to the original dimension with the help of bicubic-interpolation. These upscaled images contain lesser noise and more information compared to original noisy one and can be treated as the target Pseudo-clean images. This way, the underlying encoder-decoder based convolutional neural network, U-Net, learns to reconstruct the underlying clean image from its noisy counterpart. Loss function is proposed to include MSE (Mean Squared Error), PSNR (Peak Signal to Noise Ratio), and SSIM (Structural Similarity Measurement) terms to balance between denoising results and preservation of structural information. 5000 Images from the ILSVRC2012 ImageNet validation dataset have been selected for the training of the proposed approach. BSD300, Kodak, Set14, and SIDD datasets have been used for the purpose of evaluation. Experimental results over the datasets show that on an average the proposed method has comparable performance than that of Noise2Noise and Neighbor2Neighbor for the Gaussian, Poisson image denoising scenarios.