Deep Learning-Based Novel Image Noise Classification Model ‘MobileNoiseNet’

Image noise refers to random fluctuations in pixel intensity values caused by low light, high ISO settings, sensor limitations, compression, transmission errors, imaging sensor defects, etc., which distort the visual content and reduce the overall quality and clarity of the image. Therefore, understanding the types and sources of noise in an image is important for selecting the appropriate noise reduction techniques (filtering) and improving the overall quality of the image. Filtering is an important technique in image processing that is used to enhance the quality of images by removing unwanted noise or artifacts. However, distinguishing between diverse image noises is challenging, especially since various noise types can coexist within the same domain. For instance, MRI images might exhibit Salt and Pepper (SP), Speckle, Gaussian, Erlang, and Rayleigh noises. Therefore, the development of a mechanism to categorize these various forms of noise holds considerable significance. In this research, two deep learning-based models, MobileNoiseNet1 and MobileNoiseNet2, have been introduced. Both models are trained using the MobileNet architecture, departing from conventional transfer learning methods. Their role is to classify nine distinct types of image noise. Notably, our proposed models attain remarkable classification accuracy. In the validation set, the overall accuracy of both models stands at 94% and 92%, and this performance is further heightened in a distinct test set, achieving an impressive accuracy of 95% and 96%, respectively. This surpasses the capabilities of previous models in this domain.