Research On Transcale Space Motion Image Enhancement And Reconstruction

Some factors such as environmental illumination changes, optical or motion blur, subsampling and noise disturbance can negatively degrade the visual resolution quality of the space motion image sequences collected by multi-source vision sensors, thereby it could significantly affect the accurate estimation of location and attitude for the space moving targets and the accurate moving targets recognition and tracking, thus research on transcale enhancement and reconstruction for space motion images has great theoretical significance and application value. Although the existing enhancement and reconstruction techniques have achieved certain progress in the image processing and computer vision fields, there are still many problems to be solved in the aspects of transcale enhancement and reconstruction for space motion images. According to the characteristics of spatio-temporal and complex motion patterns for the space motion images, how to comprehensively make use of transcale correlation information in the different spatio-temporal scales of the space motion images to implement the efficient and robust transcale enhancement and reconstruction, and further improve the visual resolution and details clarity of the space motion images, propose new challenges to the space motion image enhancement and reconstruction problem. Focused on the transcale enhancement and reconstruction problem for the space motion images, this dissertation aims at studying on the key theories and technologies of the transcale feature description based on the nonsubsampled contourlet mathematical transform, the transcale self-adaptive enhancement, the spatio-temporal domain robust non-local fuzzy registration mechanism and the spatio-temporal super-resolution reconstruction based on the fuzzy registration mechanism for the space motion images. The main contributions and innovations of this dissertation are as follows:(1) Aiming at solving the problem in the existing denoising methods for the space motion images, which blurs the details of the space motion images while removing noise, this dissertation proposes a transcale adaptive space motion image denoising algorithm based on scale correlation SURE-LET (NCTSD), which further improves the space motion image edges and details retaining ability while removing noise. Based on the nonsubsampled contourlet mathematical transform, the algorithm extracts the feature information at different frequency scales and establishes the transcale description for the space motion images. A novel local adaptive threshold strategy is proposed to make determination for the coeffcients at different scales and sub-bands, and thereby the local adaptive threshold based transcale correlation calculation method is established, which further improves the distinguish ability between details and noise. The transcale adaptive denoising is implemented by solving the optimal linear combination of elementary threshold functions, which effectively retains the edge and details feature of the space motion images. Experimental results demonstrate that the proposed NCTSD algorithm gains an average promotion of19%,9%,3%and2%in term of peak signal-to-noise ratio index, and an average promotion of42%,18%,7%and5%in term of mean structural similarity index, compared to MFD, BayesShrink, BiShrink and WT_SURE-LET algorithms, respectively. Under the conditions of higher noise levels, the NCTSD algorithm obtains more obvious advantages in the denoising performance.(2) Aiming at solving the problem in the existing details enhancement methods for the space motion images, which amplifies noise while enhancing details, this dissertation proposes a transcale adaptive enhancement algorithm based on the spatio-temporal saliency (ST-CAE), which could improve the details clarity of the human eye concentrated moving target region and the algorithm time efficiency as well. We propose a spatio-temporal salient moving target region detection and extraction method based on graph theory and compressed sensing (GC-STD) to construct the spatio-temporal visual attention model. The local spatial energy is used to improve the existing bivariate statistic correlation calculation, which achieves the correlations calculation ability within and between different scales, thereby the local adaptive non-Gaussian bivariate statistic is completed. The optimal self-adaptive threshold is solved, which could effectively distinguish between noise and space image details feature. The transcale adaptive enhancement function based on spatio-temporal saliency is proposed, which could implement the details feature enhancement for the spatio-temporal salient moving targets, and effectively suppress noises. Experimental results demonstrate that the proposed ST-CAE algorithm gains an average promotion of36%and12%in term of edge energy index, an average promotion of39%and10%in term of average gradient index, and an average promotion of13%and4%in term of information entropy index, compared to existing LLDE and AUMS algorithms, respectively.(3) Aiming at solving the problem in the existing super-resolution methods for the space motion images, that is, the robustness is not strong under the complex motion patterns and the edge blur and ghosting artifacts are usually produced, this dissertation proposes a transcale super-resolution reconstruction algorithm base on Zernike moment and non-local similarity (ST-ASR). It could produce space motion image sequences with high resolution and high quality via transcale fusion of several low-resolution space motion images at different space and temporal scales. ST-ASR algorithm does not rely on accurate estimation of subpixel motion, thereby it can be adaptive to some complex motion patterns, and has higher rotation invariance effectiveness and noise robustness. An efficient iterative curvature-based interpolation method is applied to obtain the initial high resolution estimation of each low resolution motion image. The average energy based region correlation judgment and self-adaptive threshold strategies are used to improve the transcale fusion process based on Zernike moment, aiming at further improving the algorithm accuracy and time efficiency. Experimental results demonstrate that the proposed ST-ASR algorithm gains an average promotion of5%,3%and2%in term of peak signal-to-noise ratio index, and an average promotion of8%,4%and3%in term of mean structural similarity index, compared to NNI, NL-SR and ZM-SR algorithms, respectively. ST-ASR algorithm improves the time efficiency significantly while improving the reconstruction accuracy.(4) Aiming at solving the problem in the existing spatio-temporal super-resolution methods for the space motion image sequences, that is, some visual hole artifacts and details blur are easily produced in the reconstructed image, this dissertation proposes a spatio-temporal super-resolution reconstruction algorithm based on optical flow estimation and fuzzy registration mechanism (STSR), which improves the spatio-temporal resolution for the space motion image sequences at different space and temporal scales and obtains more clear and fluent space motion image sequences. The algorithm does not rely on accurate estimation of subpixel motion, and has higher rotation invariance effectiveness and noise robustness. We propose a robust optical flow motion estimation algorithm based on motion details preserving (MPOF), which makes the layered iterative optical flow calculation from coarse to fine at different resolution scales. The bi-weighted fusion idea is used to implement the spatio-temporal motion compensation for the low resolution motion image sequences. A spatio-temporal domain fast self-adaptive fuzzy registration mechanism is constructed. Experimental results demonstrate that the proposed STSR algorithm gains an average promotion of20%,8%and6%in term of peak signal-to-noise ratio index, and an average promotion of17%,7%and5%in term of mean structural similarity index, and reduces an average promotion of43%,29%and24%in term of root mean square error index, compared to POCS, NL-SR and ZM-SR algorithms, respectively. STSR algorithm not only has better Gaussian noise robustness, but also yields better performance for non-Gaussian and mixed noises (Poisson and mixed Poisson-Gaussian noise).(5) Aiming at solving the problem in the existing super-resolution methods based on the fuzzy registration mechanism, that is, the ideal super-resolution effect can not be obtained under conditions of different luminance scales, this dissertation proposes a transcale super-resolution reconstruction algorithm based on multi-source bidirectional similarity for multi-exposure motion images (ME-ASR), which implements the visual resolution enhancement for the multi-exposure space motion images by transcale fusion of the low resolution motion images at different exposure scales and different spatio-temporal scales. The algorithm can be adaptive to larger or complex motion patterns. Also it has higher rotation invariance effectiveness and is robust to noise and light. The luminance compensation for the multi-exposure motion images is completed (MBS-LC). We construct a novel fast non-local fuzzy registration mechanism based on Pseudo-Zernike moment and structural similarity. And the super-resolution reconstruction is implemented by fusion of the non-local similarity information between the motion images at different spatio-temporal scales, which could obtain motion image sequences with high dynamic range and high resolution. Experiments results demonstrate that MBS-LC algorithm can effectively reconstruct the missed details information caused by underexposure and overexposure. MBS-LC algorithm gains an average promotion of26%and17%in term of information entropy index, and an average promotion of21%and20%in term of clarity index, compared to TMP and MEFM algorithms, respectively. ME-ASR algorithm gains an average promotion of4%,5%and2%in term of peak signal-to-noise ratio index, an average promotion of12%,25%and4%in term of mean structural similarity index, an average promotion of24%,12%and11%in term of average gradient index, and an average promotion of19%,13%and11%in term of edge energy index, compared to Bicubic, Robust IBP and ZM-SR algorithms, respectively.(6) Combining the proposed NCTSD, ST-CAE, ST-ASR, STSR and ME-ASR algorithms, this dissertation designs and implements the transcale enhancement and super-resolution reconstruction system for space motion images (DIERS). DIERS system includes three logical layers:data storage layer, logical layer and user layer. The logical layer contains four function modules:denoising, detail enhancement, luminance compensation and super-resolution reconstruction. And they implements noise filtering for space motion images, detail clarity improvement for space motion targets, luminance compensation for space motion images at different exposure scales and spatio-temporal super-resolution reconstruction for space motion images at different spatio-temporal scales, respectively. Testing results demonstrate that the DIERS system could achieve better transcale enhancement and super-resolution effects, improve the visual resolution and detail clarity for space motion images in a high quality, and validate the proposed algorithms in this dissertation.