基于选择学习机制的深度图像超分辨率算法

针对深度图像传感器与彩色图像传感器的空间分辨率较差的问题,提出一种提高深度图像分辨率的算法,不同于传统方法,本文算法是基于机器学习的超分辨率选择机制,选择均值型、最大值型和中值型三种滤波器方法作为候选方法。首先运用高分辨率深度图像下采样和高分辨率彩色图像选择最优的滤波器,同时经过特征提取获得特征集；然后,高分辨率深度图像直接通过最优滤波器获得特征集；最后,这些特征集经过支持向量机(SVM)训练获得滤波器分类器。此外,本文还提出了一种频域特征向量,用于提高算法识别性能。无噪声和有噪声的深度图像实验验证了本文算法的有效性和鲁棒性,在真彩色和飞行时间深度图像的实验结果表明,本文算法的性能优于传统算法。

Learning-Based Filter Selection Scheme for Depth Image Super Resolution

Since the resolution of depth image sensors and color image sensors are very low. An algorithm is proposed to improve resolution. Unlike the traditional methods, this algorithm is based on machine learning and super resolution choosing mechanism. Choose from the mean-type, max-type and median-type filtering. Firstly, the down-sampling of high resolution depth images and high resolution color images are used to obtained the best filter and the feature sets are obtained by feature extraction. Then the feature sets are acquired by the high resolution depth images filtered by the best filter. Finally filter classifier is got from the feature sets trained by support vector machines. In addition, a new frequency-domain feature vector is designed to enhance the discriminability of the methods. The mixed sets are used in this paper to test the algorithm. Experiments on non-noise and noise images show the effectiveness and robustness. Experiments in true color and depth images of time of fight also indicate that this algorithm is better than traditional methods.