Implementation Of Blind Image Deblurring Algorithm On TMS320C6678 DSP-based Platform

Blind image deblurring is to estimate blur kernel and restore clear image under the condition of unknown blur form.It has wide application needs in medical imaging,satellite remote sensing,video monitoring and other fields.Most algorithms run on general-purpose computers,which have the disadvantages of large size and high power consumption.In some cases where the hardware environment is demanding,the embedded platform must be used.The embedded platform has the advantages of small size and low power consumption,but it also has the problem of limited computing and storage resources.This paper studies the implementation and optimization of blind image deblurring algorithm on TMS320C6678 platform.For the problem that blind image deblurring algorithm has high computational complexity and low operation efficiency,This paper proposes a method to capture the significant edge region of blurred image.We use the local area kernel estimation to replace the whole image kernel estimation.Compared with the whole image estimation,the local area estimation greatly reduces the computation and improves the performance of blur kernel estimation algorithm.For the limited storage and computing resources of the embedded platform.In this paper,we adopt targeted optimization techniques according to the characteristics of the algorithm and the hardware environment.We speed up the data transfer efficiency through EDMA3,improve algorithm execution efficiency through multi-core computing,reduce storage space requirements through array storage design,use a box filter to accelerate the window sum and so on.The experimental results show that the PSNR value of the Levin data set recovered by the algorithm implemented in this paper on the TMS320C6678 platform is above 26 dB.The running time of a single infrared image of 640*512 size was reduced from 5488.7 ms to 1213.2 ms.After optimization,77.9% of the time was saved compared with that before optimization.