基于稀疏编码的鲁棒型人脸超分辨率重建

摘要

为了减少人脸超分图像的边缘伪影和图像噪点,利用基于稀疏编码的单幅图像超分辨率重建算法,在字典学习阶段,结合L1范数引入在线字典学习方法,使字典根据当前输入图像块和上次迭代生成的字典逐列更新,得到更加精确的超完备字典对,用于图像重建.实验中进行的仿真结果表明,改进算法超分结果的峰值信噪比(PSNR)和结构相似性(SSIM)比同类型的稀疏编码超分法(SCSR)和应用在线字典学习算法的超分方法(ODLSR)均有较大幅度提升,比后者平均提升0.72 dB和0.0187.同时,视觉上有效地消除了边缘伪影,且在处理含噪人脸图像时,具备更强的去噪能力和更好的鲁棒性.%In order to reduce the artifacts and noises accompanied with the edges of face super-resolution im-ages, the improved algorithm uses the super-resolution model based on sparse coding. In the dictionary learning phase,L1-norm is combined into online dictionary learning which is used as the dictionary training method. The generated dictionary is updated column by column according to the present input image patches and the previous iterated dictionary. Thus the more accurate overcomplete dictionaries can be acquired to re-construct the final image. Comparisons of simulation results in the experiment show that the peak signal-to-noise ratio(PSNR) and structural similarity(SSIM) of the proposed method are much bigger than those of sparse coding super-resolution algorithm(SCSR) and online dictionary learning super-resolution algorithm (ODLSR). The average promotion quantity to the latter algorithms is 0. 72 dB and 0. 0187,respectively. The artifacts along the edges are eliminated effectively. The denoising capability and robustness of the pro-posed algorithm are much better than those of both SCSR and ODLSR in processing noisy face images.