车辆属性识别及跨场景标注方案

A Solution for Vehicle Attributes Recognition and Cross-Scene Annotation

提出采用一种基于深度学习的识别方法,来辅助获取训练模型中所需要的车辆属性标签。该方法首先利用海报图像构建大规模车辆属性数据集(SYSUZTE-CARS),训练基于卷积神经网络(CNN)的识别模型,再将模型迁移到监测控制场景中进行标注测试,间接获取属性标签。采用CNN+softmax分类器的结构作为基本框架,引入细粒度识别技术以进一步优化识别性能。利用5种常用车辆属性进行测试,实验结果表明:所提出的方案不仅在SYSU-ZTE-CARS数据集上的识别精度高,而且在监测控制场景下的标注结果也很可靠。

The deep learning-based solution is proposed to obtain the vehicle attributes data in recognizer training.Firstly,with poster images collected,a large scale vehicle attributes dataset-SYSU-ZTE-CARS,is built to train the convolutional neural network(CNN)-based recognition model.Then the model is moved to the monitor and control scene to mark test,and access attribute indirectly.CNN+softmax classifier is used as the basic framework,and the fine-grained identification is also used to further optimize the recognition performance.The experimental results on recognition of five different attributes show that our solution is not only with high accuracy in SYSU-ZTE-CAR dataset,but also reliable in auto-annotation under real-world surveillance.