Research Of Parallel Mechanism For GPU-Based Feature Extraction In Nonlinear Scale Space

Feature extraction algorithms in nonlinear scale space have much better robustness than ones in Gaussian scale space. However, these algorithms are very time-consuming. There were some shortcomings in the previous optimizations in Graphics Processor Unit(GPU) for these feature extraction algorithms in Gaussian scale space. These optimizations have low parallelism and cannot be implemented for feature extraction algorithms in nonlinear scale space directly.Based on the feature extraction algorithms in nonlinear scale space, a new GPU-based optimization mechanism is proposed. This optimization is used to overcome these shortcomings in previous optimizations. Firstly, based on the traditional scale space, a grouped scale space is proposed. The layers with same size are grouped into one group. In grouped scale space, all groups are independent of others. So that all groups can be proposed in parallel in GPU. Secondly, in order to eliminate the load imbalance of different sizes in grouped scale space, a data-package method is proposed. Thanks to the data-package method, not only the load imbalance can be eliminated, but also that all groups can be proposed in parallel in GPU. Finally, based on the packaged data, the scale space construction, keypoints detection and keypoints description are implemented in GPU.As the experimental result shows that, the proposed method can get 6~22x speedups with higher robustness when compared with the CPU-based feature extraction algorithms in nonlinear scale space. Even compared with the GPU-based SIFT algorithm, the proposed method can get 1.5x speedup with much higher robustness.