Noise prediction in images with VGG16 based architecture

Noise is unwanted signals added to the image during image acquisition. One of the most fundamental problems of image processing studies is noise. In order for the filter methods used to remove noise from the image to be successful, the noise type must be analyzed correctly. The types of noise that are common in the literature are gaussian, speckle and salt-pepper types of noise. It is aimed to classify these three noise types and noiseless images in the most accurate and practical way, which are added to the original images via the Matlab programming language. Convolutional Neural Networks (CNN) are frequently used in studies on images. In the study, first training and then testing were carried out for 5 different models created with CNN. ESA architectures in the study; The VGG16 network was created using transfer learning and Root Mean Square Propagation (RMSProp), Stochastic Gradient Descent (SGD), Adaptive Gradient (Adagrad), Adadelta and Adaptive Moment (Adam) optimization algorithms. It is aimed to predict the noise type and noiseless images in an accurate and practical way. If the noise prediction in the images is successful, it will be possible to use more accurate filters to remove the noise of the noisy images. The model created using the RMSProp optimization algorithm has detected a total of 12 incorrect noise types. The model created using the RMSProp optimization algorithm has the best accuracy rate with 98.75% accuracy. The accuracy rates of other optimization algorithms are 98.44% for Adam optimization algorithm, 97.29% for Adagrad optimization algorithm, 89.38% for Adadelta optimization algorithm and 57.71% for SGD optimization algorithm, respectively. With this study, it has been tried to shed light on which optimization algorithm can be preferred in order to determine the noise type in a more accurate and practical way with CNN architectures. Considering the success of CNN in image processing studies, it is thought that future studies with images can be done with more accurate and reliable filters.