oa A Novel Spatial-Domain Denoising Scheme for DoFP Polarimetric Image Sensors

Summary

In this paper, we present a novel spatial-domain denoising algorithm and directly apply it to the mosaicked Stokes sub-images generated by the division-of-focal-plane (DoFP) polarimetric image sensors. Compared to the previous implementations with the generated raw polarization images directly interpolated and demosaicked, the proposed method not only leads to significant noise reduction, but also effectively decreases the interpolated pixels' mean square error (MSE) after the interpolation process. In addition, regarding the sequence of the proposed denoising and interpolation, the polarization image quality and the interpolation MSE can be further improved by conducting the denoising before the interpolation, which has been validated by our intensive simulation results.

Motivation

Polarimetric image sensors enable a wide range of applications that are infeasible with traditional intensity/color image sensors, such as microscopy for tumor margin detection, 3-D shape reconstruction from a single image, material classification, and cancer diagnosis. By mimicking the mature Bayer-pattern-based color imaging, polarizers with different orientations are first scaled-down to micron level then mosaicked to have the full Stokes sub-images generated simultaneously in one single frame, namely division of focal plane (DoFP) (Fig. 1). However, this single-frame solution is at the expense of temporal noise and spatial resolution loss. As shown in Fig. 2, the temporal noise issue can be well-addressed by averaging multiple image frames of same micro-polarizer, which is not feasible in the aforesaid single-frame-based DoFP. In addition, interpolation algorithms are necessary to compensate the spatial resolution loss. In this paper, we propose a non-local-mean-based spatial-domain noise reduction scheme to denoise the DoFP polarization images and minimize the mean-square-error (MSE) caused by the interpolation process. Moreover, the exploration is extended to the sequence of the denoising and the interpolation as well.

Results

We compared the overall MSE of the two different sequences for different test polarization images, and it is indicated by the intensive simulation results that denoising before the interpolation can bring lower MSE [Fig. 3 (A)]. In addition, as shown in Fig. 3 (B), we compared the MSE of traditional single-frame solution, 4-frame-averaging and the proposed implementation with spatial-domain denoising, and it is indicated the proposed scheme can achieve an MSE reduction effect similar to 4-frame-averaging by spatial-domain denoising.

Acknowledgments

This work was supported by the National Natural Science Foundation of China (Grant No. 61504087), the Kongque Technology Innovation Foundation of Shenzhen (Grant No. KQCX20120807153227588), the Fundamental Research Foundation of Shenzhen (Grant No. JCYJ20140418095735624, and JCYJ20150324141711677).