基于异形卷积核的卷积神经网络图像分类方法

Based on the Heterogeneous Convolution Kernel Image Classification of Convolutional Neural Network

摘要: 卷积神经网络具有强大的图像特征学习能力，在机器学习问题中取得了突破性的进展。针对目前卷积神经网络方法中的不足之处，创新性提出了异形卷积核的概念，在不改变卷积核参数数量的情况下通过改变卷积核的形状，扩大其感受野，提高网络提取图像特征的能力。通过图像移位的方法，解决了异形卷积核可行性问题。构建了一种融合异形卷积核和矩形卷积核的卷积神经网络，使得异形卷积核能够与传统卷积神经网络有效结合。实验结果表明，相较于传统的卷积神经网络，结合异形卷积核的卷积神经网络具有更高的分类精度。

Abstract: With a strong ability to learn image features, convolution neural network has made a breakthrough in machine learning. Directing at the shortcomings of the current convolution neural network methods, the concept of heterogeneous convolution kernel is creatively proposed by changing the shape of the convolution kernel, expanding its receptive field and improving the ability of the network to extract image features, without changing the number of convolution kernel parameters. Through the method of image shift, the feasibility of heterogeneous convolution kernel is solved. The convolution neural network which combines the heterogeneous convolution kernel and the rectangular convolution kernel is constructed in order that the heterogeneous convolution kernel can be effectively combined with the traditional convolution neural network. The experimental results show that, compared with the traditional convolution neural network, the convolution neural network combined with special-shaped convolution kernel has higher classification accuracy.