A Two-stage Method for Non-extreme Value Salt-and-Pepper Noise Removal
This addresses a practical limitation in image denoising for real-world applications where noise values vary, but it is incremental as it builds on existing neural network methods.
The paper tackles the problem of removing non-extreme value salt-and-pepper noise in images, where existing methods fail because they assume noise values are exactly 0 and 255, and it proposes a two-stage CNN method that first detects noise pixels in a wider range and then denoises them, resulting in improved performance over previous approaches.
There are several previous methods based on neural network can have great performance in denoising salt and pepper noise. However, those methods are based on a hypothesis that the value of salt and pepper noise is exactly 0 and 255. It is not true in the real world. The result of those methods deviate sharply when the value is different from 0 and 255. To overcome this weakness, our method aims at designing a convolutional neural network to detect the noise pixels in a wider range of value and then a filter is used to modify pixel value to 0, which is beneficial for further filtering. Additionally, another convolutional neural network is used to conduct the denoising and restoration work.