GPU-Based Texture Mapping Research On Large-scale Realistic Scenes Rendering

The texture mapping technology based on satellite images can make geographic scene rendering more realistic, with no extra requirement of geometrical details of scenes. However, the data need to be processed in large-scale geographic scene rendering is too huge to be wholly loaded into the texture memory. On one hand, the algorithms for the compression/decompression of texture data should be researched firstly, in order to reduce the memory requirement of texture data and keep the high-quality texture details. On the other hand, the high-performance texture mapping algorithms are also required to be researched, to reduce the latency of transferring texture data and make a good use of the hardware.Based on the organization model using Tiled pyramid, several key techniques, such as the compression/decompression of texture data, the organization and scheduling of texture and geography data and the processing flow of multi-spectrum image as texture, are investigated. The main work and contributions of the dissertation are as follows.(1)The international and domestic standards of DEM data format are analyzed. The 3D geography data interposing program and the sample processing program are designed and implemented based on the interposing method in MATLAB and the sampling method by graphic hardware. The techniques of the correcting remote sensing image distortion and transforming the multi-spectrum image to the large-scale scene texture are also implemented.(2)A texture compression algorithm based on the self-growing self-organizing process and texture parallel decompression algorithms based on CUDA are proposed. In self-growing self-organizing process, the compressed texture codingbook is constructed by utilizing the method of self-growing and side associative thinking. And, after the index codes are coded by using Huffman variable length coding, the computing complexity is reduced and the quality of the compressed texture codingbook is also improved. By exploiting the flexibility and powerful computing ability of GPU, the efficacy of texture compression can be largely improved by the proposed texture parallel decompression algorithm. Experimental results show that, without loss of the reality, the required memory space for texture image and the time for transferring the texture data among the distributed network or between the internal memory and the external memory are reduced efficiently.(3)A parallel data scheduling method is proposed for multi-core processor. The latency of transferring data from external memory to texture memory can be hidden by the mechanism of constructing four-level cache, i.e., external memory, system memory, graphic memory and texture memory. Experimental results show that the method can promote the real-time functionality of scene rendering efficiently. (4)To enhance reality of geography image and promote the efficacy of texture mapping, a novel technique for dynamically generating real-time LOD texture is proposed. By utilizing the technique to sample the generated texture, the reality of scene rendering can be enhanced while the time for transferring different LOD texture data is reduced. Experimental results show that the technique can promote the efficacy of scene rendering and improve the reality of the scene.(5)A large-scale scene rendering prototype system is designed and implemented based on the above proposed techniques with the combination of the tiled pyramid model. The experiments show that the prototype system is suitable for real-time application. The various requirements of the users can be well fulfilled through the friendly human interface and the realistic sense provided by the prototype system.