Polarization Maintaining 3-D Convolutional Neural Network for Color Polarimetric Images Denoising
Hedong Liu, Xiaobo Li, Zhenzhou Cheng, Tiegen Liu, Jingsheng Zhai, Haofeng Hu
Abstract
Color polarimetric imaging could provide multi-dimensional information that provides physical properties related to object shape, surface roughness, etc. Therefore, it has been widely used in various instrumentation and measurement applications, such as 3D measurement, surface detection, and navigation. However, compared to conventional color images, polarimetric images have a lower signal-to-noise ratio and are more sensitive to noise, which leads to unpleasant noisy images and destroys polarization analysis performance. Moreover, the multiple dimensions and physical relationships among dimensions make the issue of denoise more complicated and challenging. When denoising color polarimetric images, the coherence among the space, color, and polarization is significant but has not been fully exploited by existing methods. In this paper, we propose a three-dimensional (3D) convolutional neural network for denoising color polarimetric images. 3D convolutions are applied to extract information from multiple dimensions, i.e., the space, color, and polarization. In particular, we also found that 3D features derived from color polarimetric images still satisfy the Stokes relationship after 3D convolution. In other words, polarization properties are maintained. Experiments show that the proposed denoising method can well remove noises and restore polarization information in all channels. The proposed denoising method may find important applications for other multi-dimensional imaging tasks.