深度数据结构稀疏表示理论和质量提升技术研究

High quality depth data has important applications in computer vision, pattern recognition, human computer interactions, etc. Due to the limitation of accuracy and spatial resolution of current depth cameras, the use of signal processing techniques for depth enhancement is necessary for high quality depth images. Current depth image enhancement methods using low order image models often fail to recover fine local depth details. In this project, to solve this problem, we propose a high-quality color image guided depth image reconstruction method by using the joint depth and color image sparse representation. The main research contents include: joint depth and color images structured sparse representation; joint depth and color images dictionary learning; joint sparsity-based depth image enhancement; sparsity-based depth estimation from stereo. The main contributions include: propose joint depth and color image structured sparse representation to effectively exploit the structural correlation between them; propose joint depth and color dictionary learning method to learn the structural priors of depth and color images; propose high accurate depth image enhancement method using joint depth and color sparsity model; propose sparsity-based depth estimation method from stereo images for high-quality depth estimation. In this project we expect to achieve a breakthrough in the theory of joint depth and color image sparse representation, and provide new theories and techniques for developing new high-quality depth image enhancement and super-resolution methods.

深度数据在场景三维重建、模式识别、人机交互等方面具有重大需求。由于现有深度设备精度和分辨率限制，采用信号处理方法提升深度数据精度和分辨率是获取高质量深度数据的必要手段。本课题针对深度质量提升这一难点问题，提出基于深度和彩色图像联合稀疏表示的高精度深度图像重建方法，解决现有重建方法难以充分利用高阶统计信息、难以恢复精细深度结构的问题。主要研究内容：深度和彩色图像联合字典学习、结构稀疏表示理论和模型；深度图像质量提升方法；基于稀疏模型的立体视觉深度估计方法；高精度三维场景重建方法。创新点：提出深度和彩色图像的联合字典学习和结构稀疏模型，充分挖掘它们之间的结构相关性，实现深度图像最优稀疏表示；提出基于结构稀疏正则的深度图像重建方法，以及基于稀疏模型的立体视觉深度估计方法，实现高质量深度图像重建。本项目预期在理论上有突破，技术上有创新，为实现高质量深度数据重建奠定理论和技术基础。

深度数据在场景三维重建、模式识别、人机交互等方面具有重大需求。为解决由于现有的深度设备的精度和分辨率限制而获得的图像分辨率不足的问题，本项目开展了提升深度数据精度和分辨率已获得高质量的深度数据的研究，突破了现有重建方法难以充分利用高阶统计信息、难以恢复精细深度结构的问题，提出了基于高斯尺度混合模型的图像稀疏编码模型、基于张量低秩逼近的多维图像去噪方法、基于局部结构和低秩约束的深度图像恢复方法、基于拉普拉斯尺度混合模型和非局部低秩近似的混合噪声去除方法、基于高斯尺度模型（GSM）的前景检测方法，极大的提升了图像的恢复精度，在深度图像去噪、去模糊和超分辨上获得了国内外领先的结果，共发表论文16篇，其中IEEE Transaction of image processing上发表论文5篇, International Journal of Computer Vision上发表1篇学术论文，IEEE Transaction on Pattern Analysis and Machine Intelligence上1篇，在计算机视觉领域A类国际会议IEEE International Conference on Computer Vision 2篇、NIPS上1篇。