Machine Learning for Texture Segmentation and Classification of Comic Image in SVG Compression

Abstract

Raster comic would result in bad quality while zooming in/out. Different approaches were proposed to convert comic into vector format to resolve this problem. The authors have proposed methods to vectorize comic contents to provide not only small SVG file size and rendering time, but also better perceptual quality. However, they do not process texture in the comic images. In this paper, the authors improve their previously developed system to recognize texture elements in the comic and use these texture elements to provide better compression and faster rendering time. They propose texture segmentation techniques to partition comic into texture segments and non-texture segments. Then, the <pattern> element of SVG is applied to represent texture segments. Their method uses CSG (Composite Sub-band Gradient) vector as texture descriptor and uses SVM (Support Vector Machine) to classify texture area in the comic. Then, the ACM (Active Contour Model) combining with CSG vectors is introduced to improve the segmentation accuracy on contour regions. Experiments are conducted using 150 comic images to test the proposed method. Results show that the space savings of our method is over 66% and it can utilize the reusability of SVG syntax to support comic with multiple textures. The average rendering time of the proposed method is over three times faster than the previous methods. It lets vectorized comics have higher performance to be illustrated on modern e-book devices.