Design And Implementation Of The Compressed Sensing Reconstruction Algorithm Based On FPGA

With the development of science and technology,multi-spectral camera has a great application in many fields,such as defense,medical,aerospace and aviation,etc.It is a kind of special camera which is used to shoot the image of many band at the same time.But multi-spectral camera has always been a fatal defect in technology,namely the traditional sampling method will seriously affect the time of post-processing,so that the operation speed is difficult to ascend.Because multi-spectral camera can shoot the invisible information,the public hope it can be used as a real-time monitoring/testing equipment,and the real-time performance of such facility will determine its application prospect.But the paradox is that,it is extremely difficult to implement this feature due to the sampling mode.The traditional Nyquist sampling method needs to sample the whole signal before compression.In this process,most of the sampling data will be abandoned,so why not directly access to the needed data.In 2004,Candes and Donoho proposed compressed sensing theory,which is a new theory to overcome the original sampling defects and makes the multi-spectral imaging technology has commercial and civil values.Compressed sensing theory compress the data information at the same time of perception.In other words,the collected information of CS is already the compressed data,which will greatly decrease the amount of data and fundamentally solve the shortcomings of the traditional sampling methods.The research emphasis of this system is to apply compressed sensing theory in multi spectral imaging,which is a mature imaging method.The appropriate restoration algorithm is selected to reconstruct by less data,thus realizing the multi spectral imaging.FPGA have the capability of high performance parallel computing,which makes it have the advantage in the reconstruction algorithm processing.In this system,the kernel algorithm of compressed sensing reconstruction is TwIST,namely Two-Step Shrinkage-Thresholding Algorithms.All reconstruction process is based on FPGA implementation.During the system testing phase,we assume the external scene image resolution is 256×256×15,which representing 15 different spectral bands images and each resolution is256×256.Then the scene are captured by CASSI and gray camera respectively,the obtained image resolution is 256 x 270 and 256 x 256.Finally,the amount of output data are 256×256×15,thus realizing the reconstruction of compressed sensing.In our system,the TwIST algorithm runs 200 iterations and the processing speed can up to 2.1 Ofps.In comparison to PC execution time,which is about 155.04 seconds,we gain 325.58 number of times at speed-up ration.In addition,through the comparison of PSNR between original image and the reconstruction results,the quality of the refactoring results based on FPGA can be equivalent to PC.Therefore,the system has achieved the expected function,better stability and real-time performance.