基于卷积定理的人脸验证CNN模型加速

针对人脸验证系统中复杂卷积神经网络(convolutional neural network,CNN)模型的计算负担大、运算速度慢的问题,提出使用卷积定理来加速人脸特征提取中的CNN卷积层计算,从而提升人脸验证的速度.卷积定理中,空域中的卷积运算等价于频域中的乘积运算.将耗时的卷积计算转化为频域中的乘积计算后,可能会显著减少计算量,且无精度损失.分析了用卷积定理计算卷积的时间复杂度,给出了卷积定理加速的适用条件.在进行傅里叶变换后,详细探讨了如何高效、并行地计算频域中的乘积求和,以便利用现有的并行线性代数运算库,充分发挥图形处理单元(graphics processing uni,GPU)的并行计算能力.实验结果表明:该方法对人脸验证取得了明显的加速效果,具有一定实用价值.

Accelerating CNN Models for Face Verification With Convolution Theorem

In order to alleviate the computational burden of large CNN ( convolutional neural network ) models in current face verification systems, convolution theorem was proposed, which suggested that convolution in the spatial domain was equivalent to product in the frequency domain, to speed up the convolutional layers in CNN, and consequently accelerate face verification systems. By transforming time-consuming convolutions into product operations in the frequency domain, much computation was saved without loss of accuracy. The computational complexities of convolution by using the convolution theorem and the direct computation were compared, and the conditions under which acceleration can be achieved by convolution theorem were given. After Fourier transform, the way of fulfillment of the product/sum operations in parallel was explored in detail, with the goal to fully utilize the power of GPU ( graphics processing unit ) . Results show that the proposed algorithm has achieved apparent speedups for some recent face verification models, demonstrating its effectiveness.