图像分类卷积神经网络的特征选择模型压缩方法

Convolutional neural networks model compression based on feature selection for image classification

中文关键词  卷积神经网络  图像分类  特征提取  特征选择  模型压缩

英文关键词  convolutional neural networks  image classification  feature extractor  feature selection  model compression

中文摘要

      深度卷积神经网络(convolutional neural networks, CNN)作为特征提取器(feature extractor, CNN--FE)已被广泛应用于许多领域并获得显著成功. 根据研究评测可知CNN--FE具有大量参数, 这大大限制了CNN--FE在如智能手机这样的内存有限的设备上的应用. 本文以AlexNet卷积神经网络特征提取器为研究对象, 面向图像分类问题, 在保持图像分类性能几乎不变的情况下减少CNN--FE模型参数量. 通过对AlexNet各层参数分布的详细分析, 作者发现其全连接层包含了大约99%的模型参数, 在图像分类类别较少的情况, AlexNet提取的特征存在冗余. 因此, 将CNN--FE模型压缩问题转化为深度特征选择问题, 联合考虑分类准确率和压缩率, 本文提出了一种新的基于互信息量的特征选择方法, 实现CNN--FE模型压缩. 在公开场景分类数据库以及自建的无线胶囊内窥镜(wireless capsule endoscope, WCE)气泡图片数据库上进行图像分类实验. 结果表明本文提出的CNN--FE模型压缩方法减少了约83%的AlexNet模型参数且其分类准确率几乎保持不变.

英文摘要

      Deep convolutional neural networks (CNN) feature extractor (CNN--FE) has been widely applied in many applications and achieved great success. However, evaluating shows that the CNN--FE holds abundant parameters which largely limits its applications on memory-limited platforms, such as smartphones. This study makes an effort to trim the well-known CNN--FEs, AlexNet, to reduce its parameters meanwhile the image classification performance almost remains unchanged. This task is considered as a CNN--FE model compression problem. Through carefully analyzing the parameter distribution of AlexNet, we find about 99% of parameters are in its fully connected layer but the deep features are redundant for image classification tasks with small number of categories. Moreover, we propose to convert the CNN--FE model compression problem into a feature selection problem. Specifically, a feature selection method, which is based on mutual information and a novel criterion related to the classification accuracy and the compression ratio, has been proposed. Image classification experiments on a public scene categories database and our self-built wireless capsule endoscope (WCE) bubble dataset show that our proposed CNN--FE model compression method reduces more than 83% size of the AlexNet while almost maintaining the classification accuracy.