Abstract:
The primary aim of our study is to improve the efficacy of image denoising, specifically in situations when there is a
limited availability of data, such as the BSD68 dataset. Insufficient data presents a challenge in achieving optimal outcomes due to
the complexity involved in constructing models. In order to tackle this difficulty, we provide a method that incorporates Channel
Attention, Batch Normalization, and Dropout approaches into the current REDNet framework. Our investigation indicates enhancements
in performance parameters, such as PSNR (Peak Signal to Noise Ratio) and SSIM (Structural Similarity Index), across various levels
of noise. With a noise level of 15, we obtained a Peak Signal-to-Noise Ratio (PSNR) of 34.9858 dB and a Structural Similarity Index
(SSIM) of 0.9371. At a noise level of 25, our tests yielded a PSNR of 31.7886 decibels and an SSIM of 0.8876. In addition, at a noise
level of 50, we achieved a Peak Signal-to-Noise Ratio (PSNR) of 27.9063 decibels and a Structural Similarity Index (SSIM) of 0.7754.
The incorporation of Channel Attention, Batch Normalization, and Dropout has been demonstrated to be a crucial element in enhancing
the efficacy of image denoising. The Channel Attention approach enables the model to choose and concentrate on crucial information
inside the image, while Batch Normalization and Dropout techniques provide stability and mitigate overfitting issues throughout the
training process. Our research highlights the effectiveness of these three strategies and emphasizes their integration as a novel way to
address the constraints presented by the scarcity of data in image denoising jobs. This emphasizes the significant potential in creating
dependable and effective image denoising methods when dealing with circumstances when there is a limited dataset.