基于FR-ResNet的车辆型号精细识别研究

车辆型号精细识别的关键是提取有区分性的细节特征. 以"特征重用"为核心, 以有效提取车辆图像细节特征并进行高效利用为目的, 提出了一种基于残差网络特征重用的深度卷积神经网络模型FR-ResNet (Improved ResNet focusing on feature reuse). 该网络以ResNet残差结构为基础, 分别采用多尺度输入、低层特征在高层中重用和特征图权重学习策略来实现特征重用. 多尺度输入可以防止网络过深导致性能退化以及陷入局部最优; 对各层网络部分加以不同程度的特征重用, 可以加强特征传递, 高效利用特征并降低参数规模; 在中低层网络部分采用特征图权重学习策略, 可以有效抑制冗余特征的比重. 在公开车辆数据集CompCars和StanfordCars上进行实验, 并与其他的网络模型进行比较, 实验结果表明FR-ResNet在车辆型号精细识别任务中对车辆姿态变化和复杂背景干扰等具有鲁棒性, 获得了较高的识别准确率.

Fine-Grained Car Model Recognition Based on FR-ResNet

The key of fine-grained car model recognition is to extract the discriminative feature details. Focusing on "feature reuse", aiming at efficiently extracting and using the feature details from car images, a deep convolutional neural network model named FR-ResNet (improved ResNet focusing on feature reuse) is proposed. Based on the residual structure, FR-ResNet adopts the strategy of multi-scale input, reuse of low level feature in high level network, and weight learning of feature maps to realize feature reuse. Multi-scale input can prevent performance degradation caused by too deep network and fall into local optimum. Different degrees of feature reuse for high, medium and low level network can enhance the feature transfer, efficiently reuse the features and reduce the size of parameters. The strategy of feature map weight learning applied on middle and low level network can effectively suppress the proportion of redundant features. Experimental results carried out on the state-of-the-art public vehicle datasets CompCars and StanfordCars show that, compared with other network models, FR-ResNet can obtain high recognition accuracy in car model recognition task, and is robust in the shooting angle of vehicles and the change of background.