对基于facet模型的表面检测的加速技术研究

Research on Acceleration Techniques for Facet-model-based Surface Detection

针对基于facet模型的亚体素表面检测算法计算量大的问题,提出将3维facet模型的可分滤波器递归算法与感兴趣区域加速策略相结合的加速方案。可分滤波器递归算法通过在离散正交基下将3维卷积转换为3个1维卷积并使1维卷积递归执行,使计算量与卷积核大小无关,大大节省了计算时间。采用增量算法解决了可分滤波器递归算法内存耗费量大的问题。感兴趣区域加速策略采用图像分割后提取目标的分段包围盒作为有效区域,从而大幅缩减待处理数据量。实验结果表明本文加速方案在保持原始算法精度的同时,能取得很好的加速效果。

For large computation in facet-model-based surface detection methods, an acceleration scheme combining separable filter recursive algorithm for 3D facet model with region of interest strategy is proposed. The separable filter recursive algorithm implements the 3D convolution with three 1D convolutions and allows the 1D convolution to be implemented recursively. This significantly reduces the computation time by rendering the computation independent of the kernel size. To solve the subsequent high memory consuming problem of the separable filter recursive algorithm, an incremental method is employed. As for the region of interest strategy, objects＇ piecewise bonding box extracted after image segmentation is adopted as the valid region. This can greatly decreases the amount of data to be processed. Experiment results show the presented scheme achieved excellent acceleration performance with same accuracy.