Complex convolution Kernel for deep networks
Kaizhou Li, Hong Shi, Qinghua Hu · 2016
Deep Convolutional Neural Network (CNN) is one of the most popular methods for image processing and recognition. There are many research works to improve the performance of CNNs. However, as an important part of CNNs, convolution kernel has rarely been discussed. As one Original Convolution Kernel (OCK) can only detect one type of visual feature with a fixed deformation, the networks using OCKs may learn many duplicate kernels with multiple deformations for one feature. In this paper, we propose a Complex Convolution Kernel (CCK), which can put the duplicate kernels together. Experiments on four popular datasets show the performance of networks is greatly improved by using CCKs, and suggest how to choose convolution kernels while designing networks.