A comparison of polarization image processing across different platforms

Division-of-focal-plane (DoFP) polarimeters for the visible spectrum hold the promise of being able to capture both the angle and degree of linear polarization in real-time and at high spatial resolution. These sensors are realized by monolithic integration of CCD imaging elements with metallic nanowire polarization filter arrays at the focal plane of the sensor. These sensors capture large amounts of raw polarization data and present unique computational challenges as they aim to provide polarimetric information at high spatial and temporal resolutions. The image processing pipeline in a typical DoFP polarimeter is: per-pixel calibration, interpolation of the four sub-sampled polarization pixels, Stokes parameter estimation, angle and degree of linear polarization estimation, and conversion from polarization domain to color space for display purposes. The entire image processing pipeline must operate at the same frame rate as the CCD polarization imaging sensor (40 frames per second) or higher in order to enable real-time extraction of the polarization properties from the imaged environment. To achieve the necessary frame rate, we have implemented and evaluated the image processing pipeline on three different platforms: general purpose CPU, graphics processing unit (GPU), and an embedded FPGA. The computational throughput, power consumption, precision and physical limitations of the implementations on each platform are described in detail and experimental data is provided.