改进的多标签深度学习车辆属性识别研究

为了打击假牌、套牌车及以汽车为作案工具的犯罪,且由于传统单一的车型或颜色识别已显得力不从心,因此,提出了改进的多标签深度学习车型与颜色识别模型.该模型利用卷积神经网络自主学习有用特征,利用小卷积核构建深层网络提升模型对复杂函数的表达能力,以全局平均池化取代部分全连接层,减少参数与模型所占空间内存;并利用“单模型多标签”特性将车型与颜色信息融合,使提取到的特征表现力更强.在自建数据集下的实验结果表明,该模型能获得较好的识别结果和较高的准确率,特别是对相同子品牌的不同年款的大规模车型和颜色识别效果更佳,在刑侦稽查时能有效缩小搜索范围并迅速锁定类似目标车辆信息.

Research on Vehicle Attribute Recognition Based on Improved Multi-Label Depth Learning

In order to combat false license plate vehicle and fake license plate vehicle and crimes of taking a vehicle as a tool,and because the traditional recognition model with single vehicle type or color is powerless,so the vehicle type and color recognition based on improved depth learning with multi-label is proposed.The model uses convolutional neural network to learn the useful features autonomously,and uses the small convolution kernel to construct a deep network to enhance the model expression ability to express complex functions.Then,the global average pooling is used to replace the partial fully connected layers to reduce the parameters and the model memory space.By using the “single model with multi-label” to combine vehicle type information with color information,the extracted features is more expressive.The experimental results on dataset show that the model can obtain better recognition results and higher accuracy,especially for large-scale vehicle type and color recognition of different years and styles from the same vehicle sub-brand.Therefore,it can narrow the search range effectively and lock the similar target vehicle information quickly in the criminal investigation.