COMPARATIVE STUDY OF DXT1 TEXTURE ENCODING TECHNIQUES
Jizhen Ye1, Jian Wei2, Yan Huang1, Jingliang Peng1
1School of Computer Science and Technology, Shandong University, Jinan, China
2Qualcomm Inc., San Diego, U.S.A.
In this paper, we make a comprehensive survey of many different methods to implement DXT1 (a widely used lossy texture compression algorithm). Besides that, we propose two new methods that aim for computing speed and image quality, respectively to implement DXT1 texture compression algorithm. For computing speed, we propose a new method called Lsq3d fit which achieves a very fast speed to encode texture images while keeping acceptable image quality. For image quality, we propose a new method called kmeans iteration fit and make a combination of it and the cluster fit from libsquish (an open source lib for DXTC). Kmeans iteration fit performs competitively in the quality of compressed texture images compared with the state-of-the-art DXT1 encoders, and we achieve different levels of quality by controlling the times of iteration. Finally, we test all the methods on Kodak Lossless True Color Image Suite, and CSIQ (Computational Perception and Image Quality Lab) image dataset. Our proposed methods have competitive results of speed and quality in both image datasets. The combination of cluster fit and kmeans iteration fit defeats all other methods in the quality of compressed images.