基于分类学习字典全局稀疏表示模型的图像修复算法研究

Image Inpainting Based on Non-Local Sparsity Representation with Muti-Region Learning Dictionary

基于稀疏重构的图像修复依赖于图像全局自相似性信息的利用和稀疏分解字典的选择,为此提出了基于分类学习字典全局稀疏表示模型的图像修复思路.该算法首先将图像未丢失信息聚类为具有相似几何结构的多个子区域,并分别对各个子区域用K-SVD字典学习方法得到与各子区域结构特征相适应的学习字典.然后根据图像自相似性特点构建能够描述图像块空间组织结构关系的全局稀疏最大期望值表示模型,迭代地使用该模型交替更新图像块的组织结构关系和损坏图像的估计直到修复结果趋于稳定.实验结果表明,方法对于图像的纹理细节、结构信息都能起到好的修复作用.

The performance of Sparse reconstruction for inpainting is severely dependent on the effective selection of sparse dictionary and the exploitation of self-similarity in image. For those reason, non-local sparse representation based inpainting with multi-region learning dictionary is proposed in this paper. Firstly, the available region of damaged image was clustered into regions of similar geometric structure which convey same local structural feature. The algorithm represents every region by a learning dictionary with K-SVD algorithm. The learned dictionary is then employed to recover the damaged pixel by using an expectation-maximization non-local sparse representation modal which exploits the selfsimilarity dependency in natural image to reveal the organizational structure between image blocks. Finally, the algorithm alternate the process of recovering damaged image pixel and revealing the inner organization in natural image by iterated employment of the non-local sparse representation modal until a convergent result is obtained. Experiments on texture images, structure images and real images demonstrate that the proposed algorithm is efficient and can achieve better performance.