Cartoon-Texture decomposition with patch-wise decorrelation

Cartoon-Texture decomposition (CTD) is a fundamental task and has wide applications in image processing and computer vision. To enhance separation of the cartoon and texture, existing models explicitly introduce correlation terms to decorrelate the two components. However, existing correlations usually ignore the local geometric structure information, thus insufficient to decorrelate cartoon and texture. In this work, we propose the patch-wise cosine similarity to decorrelate the cartoon and texture. The proposed decorrelation term takes the local geometric information into account and is more effective in separating cartoon and texture. Combining our decorrelation term with the regularities for cartoon (Relative Total Variation (RTV)) and texture (div(L1)-norm), we propose a new CTD model. Extended experiments show that the proposed model outperforms existing methods in CTD, especially in preserving edges of the cartoon.

Automated summary

Cartoon-Texture decomposition is a fundamental task with wide applications in image processing and computer vision. Existing models introduce correlation terms to improve the separation of cartoon and texture, but they often overlook local geometric information, resulting in insufficient decorrelation. In this work, we propose a patch-wise cosine similarity for decorrelating cartoon and texture, which takes into account the local geometric information and achieves better separation. By combining this decorrelation term with regularities for cartoon and texture, we present a new and improved Cartoon-Texture decomposition model. Experimental results show that our model outperforms existing methods, especially in preserving cartoon edges.