Image Noise Estimation with Using Pre-trained Conventional Neural Network

Noise level estimation becomes a necessary step when the noise level of the white Gaussian noise images is unknown in a realistic situation. Machine learning has brought out a series of high-performance convolutional neural networks (CNN) that work well in noise level estimation. However, studies on the pre-trained convolutional neural network are still lacking. As a result, a model based on a pre-trained Visual Geometry Group-16 network with additional regression layers was created in this research. It was then trained with a dataset that contained a series of images with different levels of additive white Gaussian noise. The results showed high performances in all ranges of the noise levels and were relatively competitive among some other algorithms. In all, a pre-trained convolutional network can estimate the noise level as accurately as some classic algorithms, and it even performed better in the high noise level range. The whole processing would be far simpler, compared to the traditional method, since CNN does not require setting parameters once the training is done.