Study on the Multi-Core DSP Parallel Processing of Polarization Images Registration

Abstract

Three-channel polarization images must be registered before pixel-level fusion processing to acquire accurate polarization characteristics information. In the condition of serial processing, the image registration efficiency is bad, and then the real-time of polarization imaging application is poor. The multi-core DSP chip which type is TMS320C6670 is selected as the polarization images processing platform. Fourier-Mellin Transform (FMT) is selected as the registration algorithm. The parallel processing of the polarization image registration is studied based on data flow model. The hierarchical task graph is designed in the parallel processing tasks partitioning. According to processing performance and functions, four DSP cores and two FFT coprocessors are divided into different processor groups in each task processing stage. Same hierarchical tasks are assigned to each processor group. According to principles including load balancing and reducing inter-processor communication, algorithms and data of each hierarchical task are assigned manually to each processing unit in the processor group. Experimental results show that the average processing time is 0.429 second while the average registration accuracy achieves 0.5 pixel, the propose parallel processing method improves the efficiency of the polarization image registration.