Joonhyun Jeong  Sung-Ho Bae

10.5626/JOK.2020.47.3.276

pointwise convolution Discrete Cosine Transform fast discrete cosine transform Convolutional Neural Network

Recently proposed light-weight neural networks maintain high accuracy in some degree with a small amount of weight parameters and low computation cost. Nevertheless, existing convolutional neural networks commonly have a lot of weight parameters from the Pointwise Convolution (1x1 convolution), which also induces a high computational cost. In this paper, we propose a new Pointwise Convolution operation with one dimensional Fast Discrete Cosine Transform (FDCT), resulting in dramatically reducing the number of learnable weight parameters and speeding up the process of computation. We propose light-weight convolutional neural networks in two specific aspects: 1) Application of DCT on the block structure and 2) Application of DCT on the hierarchy level in the CNN models. Experimental results show that our proposed method achieved the similar classification accuracy compared to the MobileNet v1 model, reducing 79.1% of the number of learnable weight parameters and 48.3% of the number of FLOPs while achieving 0.8% increase in top-1 accuracy.