A Shape-Adaptive Low-Complexity Technique for 3D Free-Viewpoint Visual Applications

作者：Chen Kuan Hung^*; Chen Cheng Hao; Chang Chih Hao; Liu Jun Yu; Su Ching Lung

来源：Circuits, Systems, and Signal Processing, 2015, 34(2): 579-604.

DOI：10.1007/s00034-014-9870-x

• Depth map generation (DMG)
• Disparity estimation (DE)
• Free-viewpoint
• Stereo matching (SM)
• Three-dimensional imaging
• View interpolation
• View synthesis (VS)

View synthesis (VS) and depth map generation (DMG) are two of the most crucial operations for facilitating the development of free-viewpoint 3D video systems. This demand for high computational complexity has motivated the development of the proposed design. Conventional designs have directly adopted dedicated hardware to increase the speed of these operations independently, which has resulted in algorithmic-level redundancies. Thus, the conventional per-pixel computation results in high redundancy in both VS and DMG operations. Accordingly, this paper proposes a shape-adaptive low-complexity (SALC) technique for removing such algorithmic-level redundancies for both VS and DMG operations. Moreover, adopting shape-adaptive searching blocks can improve the accuracy of stereo matching (SM), which subsequently enhances the resulting VS. To overcome the problem of border misalignment in VS, a novel edge alignment technique is presented for detecting and aligning misaligned borders. In addition, an unreliable region repairing (URR) technique was adopted to enhance the accuracy of DMG operations, thereby optimizing the image quality. The proposed SALC technique was verified using /Matlab simulations. Through the presented complexity reduction without any hardware acceleration, this design can achieve a high performance of 48 fps for synthesizing the virtual views of the CIF format (352 288) based on a portable platform equipped with a 1 GHz ARM Cortex-A9 RISC processor only. The experimental results revealed that the proposed SALC technique saves up to 96 % computational complexity for VS, and the computation time of DMG operations is reduced to approximately 2 % of that required for conventional SM. The average improved image quality was 32.46 dB (i.e., 7.33 % improvement from the original 30.46 dB) when using the presented edge alignment technique. When considering the overall efficiency evaluation, the proposed SALC-based SM method is 172 times more efficient than the conventional fixed block-sized SM based on the combined indices of the BPCT product. Compared with the designs of previous studies, the proposed SALC-based design obtained superior efficiency in both VS and DMG as a result of the computation reuse property.

• 出版日期2015-2