Comprehensive Approach for Image Noise Analysis: Detection, Classification, Estimation, and Denoising

Image noise is undesirable that can negatively affect the quality of digital images. It reduces the image quality and increases the processing failure ratio. It is highly recommended to remove the noise, and before removing the noise, we have to know the type of noise and estimate the parameters of noise for developing effective noise reduction techniques. This study introduces a method to effectively detect, recognize, and estimate image noise of various types (Gaussian, lognormal, Rayleigh, salt and pepper, and speckle). The proposed model consists of four stages: the first stage is detecting the noise in an image using a convolutional neural network. The second stage classifies the noisy images into one of five types of noise using a new method based on a combination of deep wavelets and support vector machines (SVM) classifier. The third stage involves estimating the parameters of the noise using maximum likelihood estimation (MLE). Finally, choosing the most suitable noise reduction technique for each type using linear and nonlinear filters and showing the capability of the suggested technique in estimating multiple noises commonly present in digital images. The proposed method utilizes a likelihood function derived from the MLE model for each noise type to estimate the noise parameters. Then used to select the most suitable noise reduction technique for each type. The quality of the denoised images is calculated utilizing the peak signal-to-noise ratio (PSNR) as the evaluation metric. The results show that the combination of wavelets with machine learning, specifically SVM, can highly enhance the results, where the accuracy was 93.043% through many experiments conducted to build a sturdy classification model. The MLE-based noise estimation method is also a reliable and accurate method for image noise estimation, especially for Gaussian, salt and pepper, lognormal, and Rayleigh noise. However, for highly noisy types such as speckle noise, further research is required to improve the estimation accuracy. This study contributes to the development of more effective noise estimation methods for improving the quality of digital images.