Polarimetric image denoising via non-local based cube matching convolutional neural network

Due to rapid advances in deep learning, many polarimetric image denoising networks have been developed and achieved promising results. However, these methods are based on general network architectures that do not fully exploit problem-specific knowledge, leading to over-smoothing results and poor generalization. Inspired by the non-local, which is an effective prior for image restoration, we propose a cube matching convolutional neural network to incorporate non-local operations into denoising models for polarimetric images. Specifically, the cube matching technique allows the denoising network to simultaneously exploit the non-local correlation and polarization relationship between the corresponding voxels of similar cubes. Rather than applying self-similarity directly in an isolated manner, the proposed cube matching module can be flexibly integrated into existing deep networks by combining with 3D convolution, achieving an effect equivalent to non-local means. This design enhances the generalization ability against various types and levels of noise. Experimental results show that using cube matching operations significantly improves denoising performance.