基于注意力引导数据增强的车型识别

车型识别在智能交通系统中发挥着重要作用。受车辆数据不足、车辆类间差异小等因素的影响，传统车型识别方法未充分利用车辆鉴别性区域的特征，导致识别准确率降低。提出一种基于注意力模块引导数据增强的车型识别方法。将ResNet-50作为骨干网络提取车辆特征，同时在网络的每个残差块后均嵌入坐标注意力模块，编码成一对方向感知和位置敏感的注意力图，以增强车辆鉴别性区域的特征表达。在此基础上，利用双线性注意力汇集操作生成增强特征图，通过对增强特征图进行注意力裁剪和注意力擦除，获取具有强鉴别性的增强数据。在Stanford Cars车辆数据集上的实验结果验证了该方法的有效性，结果表明，该方法的车型识别准确率达到94.86%，与RA-CNN、MA-CNN、WS-DAN+Inception-v3等方法相比，能够有效提高车型识别准确率和数据增强效率。

Vehicle Type Recognition Based on Attention Guided Data Augmentation

Vehicle type recognition plays an important role in intelligent transportation systems.Owing to the lack of vehicle data and small differences between vehicle classes, traditional vehicle type recognition do not make full use of the features of the vehicle discriminant area, resulting in a reduction in recognition accuracy.This study proposes a vehicle type recognition method based on attention guided data augmentation.In this method, ResNet-50 is used as the backbone network to extract vehicle features.Simultaneously, a Coordinate Attention(CA) module is embedded behind each residual block of the network to encode a pair of direction-aware and position-sensitive attention diagrams to enhance the feature representation of the vehicle discriminant area.On this basis, the bilinear attention-gathering operation is used to effectively obtain the enhanced feature image.Through attention cropping and erasure of the enhanced feature map, enhanced data with strong discrimination are obtained.The results on the Stanford Cars vehicle dataset verify the effectiveness of this method.The results showed that the accuracy of vehicle type recognition of this method reaches 94.86%.Compared with RA-CNN, MA-CNN, WS-DAN+Inception-v3, and other methods, it can effectively improve the accuracy of vehicle type recognition and efficiency of data augmentation.