图像块稀疏结构相似度邻域约束超分辨率方法
Image Super-Resolution Method with Patch Sparse Structural Similarity Neighbor Constraint
-
摘要: 在利用稀疏邻域嵌入来改进视觉伪影的图像超分辨率过程中,以稀疏特征进行高频信息估计时会产生过拟合,导致图像目标的边缘纹理部分过于平滑而丢失细节信息.针对此问题,提出图像块稀疏结构相似度邻域约束超分辨率方法.首先通过建立高/低分辨率样本块对模型并进行字典对学习,同时得到样本对模型的同构稀疏表示系数;然后以图像块稀疏结构相似度作为特征,在样本对模型中进行稀疏结构相似邻域选择;最后建立图像超分辨率目标函数,将邻域加权估计、稀疏线性组合以及下采样逼近作为约束项,采用共轭梯度算法进行模型求解.利用公共数据仿真实验和实际采集图像进行实验,并对主要参数进行验证和分析,结果表明,该方法在抑制视觉伪影的同时有效地保留了图像细节信息,峰值信噪比和结构相似度等客观评价指标也有所提高.Abstract: In image super-resolution (SR) of sparse neighbor embedding to inhibit visual artifact, the over fitting phenomenon is produced frequently in the high frequency information estimation with sparse prior. Therefore, the image details of edge and texture may be lost because of over smooth. In this paper, an image SR method is pro- posed via patch sparse structural similarity neighbor constraint. Firstly, the high-low resolution patch pair's model is constructed and the offline coupled dictionary is learned from the model. Simultaneously, we obtained the model's isomorph sparse coefficient. Secondly, we choose the image patch sparse structural similarity features to perform neighbor searching in patch pair's model. Then, the neighbor weighted approximation, the sparse linear combination and the down-sampled appro:dmation are combined as the constraint condition to construct the SR objective function. At last, the conjugate gradient algorithm is used to solve it. Some extensive experiments using common images and grabbed images are performed. The main parameters are tested and analyzed. The simulation and actual experimental results show that the method can suppress visual artifacts and preserve image detail in- formation effectively. Meanwhile, the objective evaluation results both in PSRN and SSIM are improved slightly.