基于卷积神经网络的多特征融合和图像分类

Muti-Features Fusion Based on Convolutional Neural Network for Image Classification

卷积神经网络善于挖掘图像中最具有代表性信息,但是缺少其它细节信息,而传统特征可以有效强化语义特征空间的多样性,并且不同卷积网络的语义特征相互之间存在着互补性。提出一种多特征融合的图像分类算法:根据图像颜色和梯度信息提取图像的方向梯度和颜色体积直方图(HOGCV)特征;通过预训练的网络模型提取图像的深层语义特征,并使用多种机器学习方法对不同特征进行分类训练;利用自适应加权融合算法实现多语义特征以及异构特征之间的融合。最后在数据集Cifar-10,STL-10,Cifar-100和GHIM-10K上进行验证,与单一的语义特征相比准确率提升7%~12%,与多个先进算法相比性能上有着显著的优势,证明了异构特征和多网络语义特征的互补性和自适应加权融合算法的有效性。

Convolutional neural network is good at mining the most representative information in images,but it lacks other detailed information.Traditional features can effectively enhance the diversity of semantic feature space,and the semantic features of different convolutional networks are complementary.Therefore,this paper proposes an image classification algorithm based on multi-feature fusion.Histogram of Oriented Gradient and Color Volume features of the image are extracted according to the image color and gradient information.The deep semantic features of the image are extracted through the pre-trained network model,and various machine-learning methods are used to classify and train different features.The adaptive weighted fusion algorithm is adopted to realize the fusion of multi-semantic features and heterogeneous features.Finally,it is verified on the data sets Cifar-10,STL-10,Cifar-100 and GHIM-10 K,compared with the single semantic feature,the accuracy is improved by 7%~12%,and the performance is remarkable compared with many advanced algorithms.It proves the complementarity between heterogeneous features and multi-network semantic features and the effectiveness of the adaptive weighted fusion algorithm.