Two High Speed Quantization Algorithms Luc Brun Myriam Mokhtari L.E.R.I. Reims University (I.U.T.)
Contents Quantization algorithms Our Methods Discussion
Quantization algorithms Reduce the number of colours Number of colours: 141,000Number of colours: 16
Quantization Algorithms Applications Display Compression Classification Segmentation
Quantization steps Create clusters
Quantization steps Create clusters: Squared error Partition error
Quantization steps Create clusters Compute means
Quantization steps Create clusters Compute means Create output image (inverse colormap) Quantization Inverse colormap dithtering
Type of quantization methods Three kind of Methods Top-down Bottom-up Split & Merge
Top-down methods Recursive split of the image color set
Bottom-up methods For each colour c in the image colour set Select K “empty” clusters Aggregate c to its closest cluster
Split and Merge methods Select N>K clusters (split step) Merge these clusters to obtain the K final clusters (merge step)
Our Method: Split step Create a uniform quantization.
Our Method: Merge Step Create a graph
Our Method: Merge Step Create a graph: Cluster Adjacency Graph
Our Method: Merge Step Merge of clusters: C i and C j Minimize the partition error Select i 0 and j 0 such that:
Our Method: Merge Step Merge clusters: Edge contraction
Our Method: Merge Step Merge clusters: Edge contraction
Our Method: Merge Step Merge clusters: Edge contraction
Our Method: Merge Step Merge clusters: Edge contraction
Our Method: Merge Step Merge clusters: Edge contraction
Our Method: Merge Step Merge clusters: Edge contraction
Our Method: Merge Step Merge clusters: Edge contraction
Our Method: First Inverse colormap Given a colour c Find its enclosing cluster Find its enclosing meta-cluster Map c to its mean
Our Method: Second Inverse colormap Given a color c Find its enclosing cluster Find the adjacent meta-clusters Map c to the closest mean
Our Method: Results Compared to the Top-down method [Wu-91] Image quality: First inverse colormap: slightly lower Second Inverse colormap: Improved Computing time 15 time faster Compared to the Bottom-up method [Xiang- 97] Image quality: Improved [Tremeau-96] Computing time 10 time faster
Our method: Results First inverse colormapSecond inverse colormap Wu 91Xiang 97 Original
Discussion: The idea Merge at each step the two closest clusters. Reduce the amount of data (uniform quantization) Apply an expansive heuristic: O(n 2 ) (merge step) Split & Merge strategy
Discussion: Short History Top down methods Intensively explored since 1982 [Heckbert 82] Bottom-up methods Restricted to simple Heuristics
Discussion: Short History Number of clusters Partition Error
Discussion: Short History Top down methods Bottom-up methods Split & Merge methods First attempts based on top-down algorithms.
Conclusion Possible improvements Uniform quantization Avoid empty clusters Merge Step Find a better heuristic Inverse colormap No improvement needed. Combinatorial optimisation ?
References [Wu 91] Xiaolin Wu and K. Zhang. A better tree structured vector quantizer. In Proceedings of the IEEE Data Compression Conference, pages IEEE Computer Society Press, [Xiang-97] Color Image quantization by minimizing the maximum inter-cluster distance. ACM Transactions on Graphics, 16(3): , July [Tremeau-96] A. Tremeau, E. Dinet and E. Favier. Measurement and display of color image differences based on visual attention. Journal of Imaging Science and Technology, 40(6): , 1996.IS&T/SID