Research On Compression Of Texture Images

Along with the science and technology rapid development, human has entered a new digital era. Virtual reality technology has brought people a completely new immersive experience. In order to provide users better virtual reality experiences, the virtual pictures must have enough sense of reality, and support user’s real-time interaction. So interactive real-time photo-realistic rendering become a hot research area. Texture mapping is the technique that is used to balance real-time computing and reality. It has been widely used since proposed. Due to the limited bandwidth and memory storage size, it is difficult to render many high resolution texture images real time so that texture compression is introduced. Texture compression can not only improve the texture cache utilization rate, but also reduce the system data transmission pressure effectively. To a large extent, it solves the problem of real-time photo-realistic rendering and has a profound influence of texture compression field.As for the particularity of texture image compression, we can’t merely consider the compression rate and the quality of decompression texture image, must also consider many other factors. Such as, is the algorithm compatible with the mainstream graphic cards, the hardware implementation difficulty, efficiency, cost and so on. For this reason, we can’t do research as we wish. We must follow the existing standard and framework of texture compression. In this paper, we choose to improve the most popular algorithm used in windows desktop system DXTC. We focus on DXT1 encoder, as other kinds of encoders are based on it. Then we make a comprehensive research on existing DXT1 encoding methods, including some open source library and some methods that are integrated into compression tools. Based on this, we propose novel methods to implement DXT1, which include two main algorithms. One is a new method called Lsq3dfit which achieves a very quick speed to encode textures, the other one is called k-means iteration fit. Experiment results show that our fast encode method beats all other methods, quality method are suitable for blocks that has a high degree linearity of color points. The combination of our k-means iteration fit and cluster fit defeats all other methods in all the quality assessment metrics.