基于区域自适应学习的人脸图像超分辨率复原

Region Adaptive Learning-Based Super Resolution for Human Face Images

提出一种基于区域自适应学习的人脸图像超分辨率复原算法。算法根据图像的纹理特征将人脸分为平坦区和细节区。对面部平坦区直接采用双线性插值放大;对于眼睛、鼻子和嘴等细节区,采用分类预测器重建高频信息。在细节区,将相似纹理结构的图像块分为一类,对每类纹理结构分别训练线性预测器,进行高频信息预测。实验结果表明,本算法在图像重建质量的主观效果和实现速度上都有很好的表现。

A novel region adaptive learning-based super resolution algorithm for human face posed, which divides a face image into fiat regions and detailed regions. Flat regions are magnifi images is proed using bilinear interpolation; and the detailed regions, such as eyes, mouth and nose, are super resolved using classified predictors. The local patches within the detailed regions are classified into several categories according to the orientation of texture structure. A linear predictor is trained up for each category to infer the high-frequency component. Experimental results show that both the visual quality and the computational cost are improved.