基于语义DCNN特征融合的细粒度车型识别模型

针对深度卷积神经网络模型缺乏对语义信息的表征能力,而细粒度视觉识别中种类间视觉差异微小且多集中在关键的语义部位的问题,提出基于语义信息融合的深度卷积神经网络模型及细粒度车型识别模型.该模型由定位网络和识别网络组成,通过定位网络FasterRCNN获取车辆目标及各语义部件的具体位置;借助识别网络提取目标车辆及各语义部件的特征,再使用小核卷积实现特征拼接和融合;最后经过深层神经网络得到最终识别结果.实验结果表明,文中模型在斯坦福BMW-10数据集的识别准确率为78.74%,高于VGG网络13.39%;在斯坦福cars-197数据集的识别准确率为85.94%,其迁移学习模型在BMVC car-types数据集的识别准确率为98.27%,比该数据集目前最好的识别效果提高3.77%;该模型避免了细粒度车型识别对于车辆目标及语义部件位置的依赖,并具有较高的识别准确率及通用性.

Fine-Grained Car Recognition Model Based on Semantic DCNN Features Fusion

As the deep convolution neural networks (DCNN) lack the ability of representation of semantic information,while visual differences between classes are small and concentrated on key semantic parts during the fine-grained categorization,this paper proposes a model based on fusing semantic information of DCNN features,which is composed of the detection sub-network and the classification sub-network.Firstly,by use of the detection sub-network we capture the definite position of car object and each semantic parts through Faster RCNN.Secondly,the classification sub-network extracts the overall car object features and semantic parts features of the object via DCNN,then processes the joint and fusion of features by using small kernel convolution.Finally,we obtain final recognition result through deep neural network.The recognition accuracy of our model is 78.74% in Stanford BMW-10 dataset,which is 13.39% higher than the VGG network method and 85.94% in the Stanford cars-197 dataset.And the recognition accuracy of the transfer learning models in BMVC car-types dataset is 98.27%,which is 3.77% higher than the best recognition result of the dataset.Experimental results show that our model avoids the dependence of the fine-grained car recognition on the positions of car object and semantic parts,with high recognition accuracy and versatility.