一种基于深度残差网络的车型识别方法

A Vehicle Model Recognition Algorithm Based on Deep Residual Network

针对传统的车型识别方法提取的特征的可分性较差、鲁棒性不足等问题,提出一种基于深度残差网络的车型识别方法。相比于传统的特征提取方法,深层网络模型具有模型参数更为充分完善的优势,同时也更加适合于处理大规模的数据集,其提取的特征具有天然的层次结构,类型也更加丰富。深度残差网络使用的残差单元可以改善深层网络模型寻优的过程,减少模型收敛的时间开销。在深度残差网络的基础上添加类别中心正则化的约束可以改善特征的分布空间,强化同一类别内的特征的相似性及不同类别的特征的可区分性,进一步提高模型的分类性能。训练时,将训练过程分为两个步骤,分别使用不同的数据集进行训练可以提高训练的效率,充分利用预训练模型的优势。实验结果表明,该算法在识别精度上优于传统的车型识别方法。

In view of the poor differentiability and robustness of the feathers extracted from traditional vehicle recognition method, we propose a useful method based on deep residual network for vehicle recognition. Compared with traditional methods of feature extraction, deep layer network model has the advantages of model parameters with more full improvement and is also more suitable for processing large datasets. The extracted features are also hierarchical and much more abundant. Deep residual network contributes to optimizing the deep model with the residual learning module, speeding up the computation progress. Attaching the center loss task on deep residual net- work will improve the distribution of features, and strengthen the similarity of features attach to the sample genre and the discrimination between features attach to different genres, so us to further advance the performance of classification. At the same time, the training progress can be divided into two steps with different datasets respectively, which can improve the efficiency and make full use of the advantages of pre-trained models. The experiments show that the proposed algorithm is superior to the traditional methods in classification.