Abstract: In order to improve the accuracy of sparse stack coding for vehicle type identification, this paper proposes a vehicle identification method based on improved sparse stack coding. The layer-by-layer unsupervised method is used to train the network structure, and the feature dictionary is learned from a large number of unmarked data sets. Then, the convolution and pooling modules are introduced on the basis of sparse stack coding, and the learned feature dictionary is taken as a convolution kernel. Feature map of the image is obtained by convolving and pooling the image containing the vehicle. Finally, supervised fine-tuning is performed on a small number of tag data sets by using the softmax classifier. By experimenting on the BIT-Vehicle dataset, the improved algorithm is superior to the traditional sparse stacking algorithm. In the dataset with less labeling, the recognition accuracy is better than the neural network algorithm.

Key words: vehicle type classification, sparse stack coding, convolution, pooling, feature dictionary

摘要： 为了提高稀疏栈式编码对车型识别确率，提出了一种基于改进稀疏栈式编码的车型识别方法。使用逐层无监督方法来训练网络结构，并从大量的无标记的数据集中学习得到特征字典，在稀疏栈式编码的基础上引入卷积和池化模块，把学习得到的特征字典作为卷积核，通过对含有车辆的图像进行卷积和池化操作获得图像的特征图；最后通过使用softmax分类器在少量标签数据集上进行有监督的微调。在BIT-Vehicle数据集上的实验结果表明，改进后的算法优于传统稀疏栈式编码算法，在标注较少的数据集中，识别的准确率优于神经网络算法。

关键词: 车型识别, 稀疏栈式编码, 卷积, 池化, 特征字典