基于梯度域计算模型的彩色图像去运动模糊与特征提取方法研究

Motion deblurring of color images and feature extraction are two difficulties in the fields of image processing and recognition and have attracted a lot of attention in the scientific community. Restoration results by the existing motion deblurring methods often appear over-smooth, removing edges and textures which are important for visual realism. For data corrupted by outliers, such as occlusions, illumination changes, and cast shadows, features extracted by the existing feature extraction methods can be arbitrarily biased. Previous studies have shown that image gradient provides useful information and gradient domain computing models provide significant technical support for image processing. On one hand, the heavy-tailed gradient distribution is important prior information in image restoration. On the other hand, features will be more robust to contour noise by means of gradient vector decomposing technique. The main aim of this project is to research affect mechanism of gradient information to the restoration resluts of motion-blurred color images and feature extraction. The main research contents of this project include the following aspects. Firstly, nondirectional difference operators are used to compute directional difference operator, based on which, a new gradient estimator of color images that can get accurate estimation of both gradient magnitude and gradient orientation, is modelled, which will improve denoising effect of color images. Secondly, a cumulative penalty function is defined based on convergence problem of empirical gradient distribution to a reference distribution and a gradient distribution matching model is established. A new motion-blurred image restoration method is proposed based on the model, which can provide abundant textures and edges. Thirdly, correlation computing model is established based on gradient orientations and cosine function. A new feature extraction method, which is robust to outliers is proposed based on the correlation computing model. The research will provide theoretical foundation for the consummating of gradient domain computing models and for the technical development in the fields of motion-blurred color image restoration and feature extraction, which is significant in academic aspect and application prospect.

彩色图像去运动模糊与特征提取是图像处理和识别领域的两个难点，令人关注。当前研究存在运动模糊图像恢复结果的纹理过于平滑、对遮挡等异常破坏的特征提取效果较差等问题。前期工作表明，梯度信息为图像恢复提供了重要的先验知识，通过梯度矢量分解技术能够提高特征对轮廓噪声的鲁棒性。本项目将研究梯度信息对彩色图像去运动模糊和特征提取的作用机制，拟基于无方向差分算子获取方向差分算子，建立同时能估计梯度大小和方向的彩色图像梯度估计模型，提高对噪声的鲁棒性；基于此，从经验梯度分布对参考梯度分布的收敛性角度构造累积代价函数，建立基于该函数的梯度匹配模型，构造能丰富纹理信息的彩色图像去运动模糊算法；采用梯度方向差分与余弦函数相结合的方法，建立相关性计算模型，构造对异常破坏鲁棒的特征提取算法。该研究将为梯度域计算模型的完善以及彩色图像去运动模糊和特征提取领域的技术发展提供重要的理论依据，具有重要的学术意义和应用前景。

彩色图像去噪、图像去运动模糊、特征提取是图像处理和识别领域的研究难点，令人关注。本课题对这几个问题进行了研究。.对于彩色图像去噪问题，针对全变分去噪模型易受到梯度的影响而且常常会丢失图像细节信息的缺点，提出了自适应权重全变分图像去噪模型来恢复损坏的加性高斯噪声图像。该方法在去除噪声的同时能较好的保持图像纹理和边缘。提出了一种基于引导滤波与游长概率统计的图像去噪方法，并将该方法用在古代碑文图像去噪中。该方法不仅能抑制图像噪声，而且能够保护汉字的结构。.对于图像去运动模糊问题，提出了一种基于 L1正则化算法的模糊核估计方法。该方法根据估计的模糊核恢复图像边缘信息，并对恢复后的图像继续估计模糊核，两步交替计算直到求得正确的模糊核。该方法不仅具有一定的抗噪性，而且能提高模糊核估计的精度。为了提高模糊核估计的效率，提出了基于灰度共生矩阵提取丰富边缘区域的模糊复原方法。该方法依据灰度共生矩阵的信息选取出包含边缘信息相对较多的区域，用提取出来的丰富边缘区域代替整幅运动模糊图像进行模糊核的估计。针对运动模糊图像复原中的振铃效应问题，提出一种正弦积分拟合的图像复原边界振铃效应抑制方法。该方法在对待处理的模糊图像进行边缘延展的基础上，分别利用正弦函数积分方法和双正弦函数积分方法对单向过渡区域和双向过渡区域进行窗函数计算。.对于特征提取，研究了受眼镜遮挡的面部不变性特征提取问题，提出了一种基于边缘信息投影定位镜架横梁进行人脸图像上眼镜特征提取的方法，该方法能用于人脸识别等领域。为了提高特征提取的效率，提出基于有效集算法优化极限学习机模型，并将梯度直方图与该优化极限学习机模型相结合进行人体特征提取，用于人体检测。该方法减少了人体检测中训练过程的计算代价，提高了特征提取效率。.该研究将为图像去噪、图像去运动模糊和特征提取领域的技术发展提供重要的理论依据，具有重要的学术意义和应用前景。