Influence of Kernel Size and Layer Size for MAE Score and Model Fit in Image Denoising

The presence of noise in images is one of the challenges that complicate image processing. Numerous methods have been proposed to mitigate the impact of noise on denoising image performance. However, the application of CNN for denoising image performance optimization in the presence of noise is still a new area of research. In this paper, the configuration of kernel size and number of layers will be investigated in image denoising using CNN (Convolutional Neural Networks), specifically, in U-Net architecture. The images used are from BSD500 which consists of 500 natural images. Then, the hyperparameters will be evaluated in terms of many factors, such as mean absolute error (MAE) score and model fit to show its influence. The results show that the average score for MAE is 0.11 based on the model and variation of both kernel size and layer size, with the highest score found at kernel size=2 and layer size=64. The results show that to get a fit model, the experience should use epoch number 10. It shows that the influence of kernel size has a negative influence on MAE whereas layer size has a beneficial effect on model fit. However, there is an optimal point beyond which denoising results can decrease. To achieve better results, future researchers can compare it with other denoising models to ensure our findings.