论文部分内容阅读
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.
As a feature extractor (CNN-FE), convolutional neural networks (CNNs) have been widely used in many fields and have been significantly successful. According to the research evaluation, CNN-FE has a large number of parameters, which greatly limits CNN-FE is applied to devices with limited memory such as smart phones.In this paper, we use the AlexNet convolution neural network feature extractor as the research object, face the image classification problem, and reduce the CNN while keeping the image classification performance almost unchanged -FE model parameters.A detailed analysis of the parameter distribution of AlexNet layers reveals that the fully connected layer contains approximately 99% of the model parameters, and there are redundancies in the features extracted by AlexNet when there are fewer image classification categories. , The CNN-FE model compression problem is transformed into a deep feature selection problem, and a new feature selection method based on mutual information amount is proposed to reduce the CNN-FE model compression rate by jointly considering the classification accuracy and the compression ratio.In the open scenario Classification database and self-built wireless capsule endoscope (WCE) bubble image database.Results It shows that the proposed CNN-FE model compression method reduces the AlexNet model parameters by about 83% and the classification accuracy remains almost unchanged.