半径间隔界驱动卷积神经网络模型的图像识别

Image Recognition Via Radius Margin Bound Guided CNN Model

基于支持向量机(SVM)的卷积神经网络(CNN)模型结合了大间隔原理,在图像识别中表现出了优异的泛化性能。然而,该方法忽视了一个关键:SVM的泛化性能不仅取决于不同类之间的间隔,还与所有样本的最小包含球(MEB)的半径有关。针对这一事实,文章提出一种基于半径间隔界(RMB)驱动的CNN模型的图像特征提取和识别的方法。与传统CNN模型相比,该模型采用基于SVM泛化误差界的策略来指导CNN深度模型学习和相应分类器构建,不仅考虑了不同类别之间的间隔,还考虑了MEB的半径。该模型能提高深度卷积模型的泛化能力而不会额外增加网络的复杂度,还能够应用于不同的深度模型中而不受限于某一特定的网络结构。在多个数据集上的实验结果表明,相比于基于Sofmax损失的CNN模型、基于中心损失的CNN模型以及基于SVM的CNN模型,该模型能够提取到鉴别性更强的图像特征,取得更高的识别率。

The convolutional neural network(CNN)model based on support vector machine(SVM)combines the margin maximization principle and achieves excellent generalization ability in image recognition applications.However,this method ignores a key fact that the generalization ability of SVM depends not only on the margin between the different categories,but also on the radius of the minimum enclosing ball which contains all the samples.Aiming at this problem,a CNN model driven by radius margin bound(RMB)is proposed to extract and identify the image features.Compared with the traditional CNN models,the proposed method not only considers the margin between different categories of the image features,but also further considers the radius of the minimum enclosing ball.In essence,the proposed CNN model adopts a strategy, which based on SVM generalization error bound, to guide the learning of the CNN modeland the construction of the corresponding classifier. The model can improve the generalization ability of thedeep convolution model without adding additional network complexity, and can also be applied to differentdepth models without being limited to a particular network structure. The experimental results on multipledatasets show that compared with the CNN model based on Sofmax loss, the CNN model based on centerloss and the CNN model based on SVM loss, the proposed method can extract image features that havemore discriminative power and further obtain higher recognition rate.