Coherent Ladar Image Restoration By Non-local Filtering Algorithms

Digital image processing has played an important role in ladar imaging. It is howto remove the speckle noise in intensity images and range abnormality noise in rangeimages effectively that has an directly effect on the precision and performance of entireladar system. Non-local means filtering (NLM), block-matching and3D filtering(BM3D) are newly proposed denoising algorithms which based on non-local filtering inrecent years. Especially is the later getting better results in restoring images withGaussian noise. Due to the good performace of these algorithms and the purpose ofimproving the level of ladar image restoration, several image ratoration techniquesbased on non-local filtering algorithms are studied with ladar intensity images and rangeimages in this paper.To find out a reasonable quality metrics of ladar intenstiy images and range images,structural similarity, peak signal noise ratio and speckle index are utilized to evaluateintensity images, and structural similarity and root mean square error are utilized toevaluate range images. The evaluation results are analyzed and discussed.Several non-local filtering algorithms, which include NLM and its improvingforms proposed by other experts, patch-based locally optimal Wiener filter and BM3D,are studied with simulated intensity images which have diferent Carrier to Noise Ratio(CNR) or some targets with different CNR. According to the multiplicativecharacteristic of speckle noise in intensity images, homomorphism transform is addedand parameters in algorithms are properly chosen. The filtering results are analyzed anddiscussed on the homogenization of target gray values, the sharpness of target edges,quality metrics and so on.Due to the characteristic of preserving fine details in NLM, and integrating withmedian filter and background suppression, the image fusion denoising algorithm (IFD)is proposed. According to the property of range abnormality noise, which is a uniformdistribution over the entire range interval, the non-local probability statistics filteringwhich can preserve true range data is proposed. Several non-local algorithms, includingboth of the proposed algorithms, NLM and BM3D, are studied with range images ofdifferent CNR. To increase the relevance with reality, a complex scene model withtargets having different CNR is set up. The range images of the model are denoised by IFD. All results are analyzed and discussed.To verify the correctness of simulation, real intensity images and range images ofladar are processed by non-local algorithms used above.