摘要： 卷积神经网络因缺乏空间不变性造成分类精度不高，且由于复杂度过高导致分类效率较低。提出一种利用空间变换网络和异构卷积滤波器的SH_ImAlexNet网络，应用于肝硬化样本识别。改进卷积神经网络AlexNet的结构和参数以满足肝硬化样本尺度要求，引入空间变换网络层增强特征提取能力与空间不变性，采用异构卷积滤波器替换部分卷积核降低复杂度并提升鲁棒性。实验结果表明，该网络的分类效果较AlexNet、VGG等传统网络更优，在小样本数据集和大样本数据集上的识别率分别达到98.28%和95.67%，空间复杂度和时间复杂度更低且运行效率更高。

关键词: 空间变换网络, 异构卷积滤波器, AlexNet模型, 卷积神经网络, 肝硬化识别

Abstract: The lack of space invariance of Convolutional Neural Network(CNN) results in low classification accuracy and low classification efficiency due to its high complexity. This paper proposes a SH_ImAlexNet network based on Spatial Transformer Network(STN) and Heterogeneous Convolution(HetConv) filter for the identification of liver cirrhosis samples. The structure and parameters of CNN AlexNet are optimized to fit into the size of liver cirrhosis samples, and the STN layer is introduced to enhance the feature extraction ability and spatial invariance. The HetConv filter is used to replace part of the convolution kernel to reduce the complexity and improve the robustness. Experimental results show that the classification performance of the proposed network is better than that of traditional networks such as AlexNet and VGG. The recognition rate of the network on small sample dataset and large sample dataset reaches 98.28% and 95.67% respectively. It provides a higher operation efficiency with reduced space complexity and time complexity.

Key words: Spatial Transformer Network(STN), Heterogeneous Convolution(HetConv) filter, AlexNet model, Convolutional Neural Network(CNN), identification of liver cirrhosis

中图分类号: 

• TP183