Vehicle Logo Recognition with Small Sample Problem in Complex Scene Based on Data Augmentation
Vehicle Logo Recognition with Small Sample Problem in Complex Scene Based on Data Augmentation
复制标题
基于数据增强的复杂场景小样本车标识别问题
DOI：
10.1155/2020/6591873
复制
发表时间：
2020-07-09
影响因子：
--
通讯作者：
Du, Pengqiang
中科院分区：
文献类型：
--
作者：
Ke, Xiao;Du, Pengqiang
车辆自动识别是智能交通系统（ITS）领域的一个重要课题，而车辆标志是车辆最重要的特征之一。因此，车辆标志检测与识别是重要的研究课题。针对车辆标识面积过小无法检测、数据集过小无法对复杂场景进行训练等问题，考虑到识别速度和对复杂场景的鲁棒性，采用基于数据优化的深度学习方法对复杂场景下的车辆标识进行识别。提出了三种汽车标志数据增强策略：交叉滑动分割法、小帧分割法和高斯分布分割法。针对样本量小的问题，我们采用交叉滑动分割的方法，在不改变原始车辆标志图像长宽比的情况下，有效地增加了数据量。为了扩大图像中标识的面积，我们开发了小帧方法，提高了小面积车辆标识的检测效果。为了丰富车辆标志在图像中的位置多样性，提出了高斯分布分割方法，结果表明该方法是非常有效的。我们的方法在YOLO框架下的F1值为0.7765，精度大大提高到0.9295。在Faster R-CNN框架下，我们的方法的F1值为0.7799，也比以前好。实验结果表明，与传统方法相比，上述优化方法能更好地表征车辆标志的特征，实验结果得到了改善。
Automatic identification for vehicles is an important topic in the field of Intelligent Transportation Systems (ITS), and the vehicle logo is one of the most important characteristics of a vehicle. Therefore, vehicle logo detection and recognition are important research topics. Because of the problems that the area of a vehicle logo is too small to be detected and the dataset is too small to train for complex scenes, considering the speed of recognition and the robustness to complex scenes, we use deep learning methods which are based on data optimization for vehicle logo in complex scenes. We propose three augmentation strategies for vehicle logo data: cross-sliding segmentation method, small frame method, and Gaussian Distribution Segmentation method. For the problem of small sample size, we use cross-sliding segmentation method, which can effectively increase the amount of data without changing the aspect ratio of the original vehicle logo image. To expand the area of the logos in the images, we develop the small frame method which improves the detection results of the small area vehicle logos. In order to enrich the position diversity of vehicle logo in the image, we propose Gaussian Distribution Segmentation method, and the result shows that this method is very effective. The F1 value of our method in the YOLO framework is 0.7765, and the precision is greatly improved to 0.9295. In the Faster R-CNN framework, the F1 value of our method is 0.7799, which is also better than before. The results of experiments show that the above optimization methods can better represent the features of the vehicle logos than the traditional method, and the experimental results have been improved.