A Hybrid Filter Training and Design Method for Adaptive Noise Cancellation

A hybrid filter training and design method for adaptive image noise cancellation with establishing noise channel model is presented in this paper. Trainings are performed with input and output test images for channels of different types and/or different intensities to establish channel-adaptive hybrid filter models. In practice the image transferred through the specific channel is filtered blindly by corresponding model to maintain detail and eliminate noise simultaneously. For certificating the efficiency of this design method, Gaussian weighted median filters are adopted to remove well known channel noise, i.e., pepper and salt noise and uniform distributed impulse noise, in this framework. The results outperform some prior methods markedly. This design method can be generalized to other filters with adaptability to treat different channel noise flexibly.