Sparse Coding Way For Reconnaissance Image Compression

Two sparse coding methods for reconnaissance image compression,Set Partition Coding(SPC)and Compression Sensing(CS)are studied in this paper.The paper focuses on versatility and lossless encoding efficiency of SPC,decoding quality of CS,to present image coding methods systematically for characteristic reconnaissance platforms.The major contributions are outlined as follows.(1)To solve the problem of current SPC methods lacking versatility,a general SPC method named SPACS has been constructed.The existing Tree-SPC(T-SPC)methods and Block-SPC(B-SPC)methods are focused on exploitation of interband correlation and intraband correlation of wavelet transformed images,respectively.General Tree(GT),designed in our paper,has ability to represent tree set and block set,simultaneously.With help of normalized GT partition rules,the general SPC method named SPACS(k,p)was constructed.Partition mode and initial form of GTs are decided by parameters k and p.This 2 flexible parameters k and p introduce versatility to SPACS: SPACS trends to T-SPC method as k and p decrease,exploits more interband correlation and suits to compressions of images with few high component,and SPACS trends to B-SPC method as k and p increase,exploits more intraband correlation and suits to compressions of images with more high component.The paper presents concepts of amplitude bit encoding coefficient amplitude directly,location bit and unnecessary bit encoding coefficient coordinate indirectly,to quantitatively analyze coding performance of SPACS by count the number of the 3 different types of bits.The numerical experiments for reconnaissance images with various statistical properties,show that,for compression of each reconnaissance image,the unnecessary bit output from SPACS decreases by 4.1% on average,and the mean PSNR increases by 0.18 dB.SPACS is a general coding method and is superior to previous SPC methods,which has solve the universality problem for compressions of various reconnaissance images.(2)To increase the lossless coding performance of SPACS,SPACS_C has been constructed.Some mathematical expressions were established to describe amounts of different types of bits used in encoding some special GTs in SPACS.The conclusion deduced from the expressions is that the lossless coding performance of SPACS could be improved if the location bit1 was reduced.The previous SPC methods process all GTs with only one coding way named“test&partition”.SPACS_C,which is SPACS equipped with a classifier,calculates the bit costs of two different coding way named “test&partition” and “partition&test” for a GT,then chooses the coding way with fewer bit cost to process the GT.The process mode used in SPACS_C suits to the data characteristics of in wavelet transformed images that sparseness of bottom bitplanes decrease rapidly.According to the results of various reconnaissance images tested by SPACS_C,the lossless encoding performance of SPACS_C was improved and superior to that of JPEG2000 and JPEG-LS.The coding performance of SPACS_C satisfy the high quality requirement for image encoded losslessly.(3)To improve the decoding quality of image CS method with very lower encoding complexity,SPL-TV algorithm has been constructed.The objective function of Smoothed Projected Landweber(SPL)algorithm was imported with total variation(TV)of image,which characterizes the structure of the image,to build optimization problem for Smoothed Projected Landweber regularized with Total Variation(SPL-TV).Hard threshold filtering and project onto convex sets were employed to approach the partial derivatives of norm-terms in the objective function of SPL-TV algorithm,then the optimization problem of SPL-TV algorithm was solved using gradient descent.The numerical result of visible aerial images show that SPL-TV algorithm suppresses the decay of decoding performance appear in iterative end stage of SPL algorithm.There is a obvious improvement for visual effect of SPL-TV decoding images,and the mean PSNR corresponding to decoding images of SPL-TV algorithm increases 1.17 dB.To improve decoding performance of CS method,SPL-TV algorithm exploits the local smoothness and bounded variations of images synchronously,and SPL-TV algorithm suits to small reconnaissance platforms adopting encoding method with very low complexity.