MCVSVM驱动卷积神经网络模型的图像识别

Image Recognition Via MCVSVM Guided CNN Model

针对支持向量机(SVM)驱动的积神经网络(CNN)模忽视特征的空间分布信息对模型泛化性能的不足,提出一种最小类内方差支持向量机(MCVSVM)驱动的CNN模型来处理图像识别任务。得益于MCVSVM中类内散度矩阵的引入,提出的CNN模型不仅考虑异类图像特征间的间隔,同时能够利用特征空间中特征向量的分布信息对CNN进行微调。在五个大规模数据集上的实验结果表明,相对于SVM驱动的CNN模型,MCVSVM驱动的CNN在实验数据集上的Top-1识别准确率最大提高4.44%。MCVSVM驱动的CNN具有更强的泛化能力以及更高的识别准确率。

Aim at the problem that The Convolutional Neural Network(CNN)module driven by Support Vector Machine(SVM)ignores the influence of the spatial distributing information of feature vectors on the generalization performance of the model.A CNN model driven by Minimum Class Variance Support Vector Machines(MCVSVM)is proposed to handle image recognition tasks.Thanks to the introduction of the within-class scatter matrix in MCVSVM,the proposed CNN model not only considers the within-class margins,but also uses the distribution information of feature vectors in the feature space to fine-tune the CNN.Experimental results on five large-scale data sets show that Compared with the SVM-driven CNN model,the top-1 recognition accuracy of the MCVSVM-driven CNN on the experimental data set increase to nearly 4.44%.MCVSVM-driven CNN has stronger generalization ability and higher recognition accuracy.