雾霾环境下基于PLATE-YOLO的车牌检测方法

License Plate Detection Method Based on PLATE-YOLO in Smog Environment

针对目前车牌识别领域中,雾霾环境下车牌检测准确率低的问题,本文提出一种基于深度学习的抗雾霾车牌检测方法,该方法能够检测民用车牌和机场民航车辆车牌。该方法首先利用一种基于卷积神经网络的去雾算法对车牌图片进行去雾预处理,然后将处理过的无雾霾图片送入PLATE-YOLO网络中检测车牌的位置。该PLATE-YOLO网络是本文针对车牌检测的特点,对YOLOv3网络做了修改后得到的适用于车牌检测的网络。主要改进点有两处:第一,提出了一种基于层次聚类算法的锚盒(Anchor Box)个数和初始簇中心的计算方法;第二,针对车牌目标较大的特点,对网络的多尺度特征融合做了优化。优化后的PLATE-YOLO网络更适合于车牌检测,且提高了检测速度。实验证明,PLATE-YOLO网络检测车牌的速度较YOLOv3提高了5 FPS;在雾霾环境下,经去雾预处理的PLATE-YOLO车牌检测方法比未经去雾处理的车牌检测方法准确率提高了9.2%。

Aiming at solving the problem of low license plate detection accuracy in the smog environment in the current license plate recognition field,a deep learning-based anti-smog license plate detection method is proposed in this paper.The method can detect civil license plates and airport civil aviation vehicle license plates.Firstly,a deconvolution algorithm based on convolutional neural network is used in the method to defog the license plate image.Then,the processed smog-free image is sent to the PLATE-YOLO network to detect the position of the license plate.The PLATE-YOLO network is a network for license plate detection that has been modified for the YOLOv3 network.The PLATE-YOLO network is suitable for the license plate detection network.There are two improvement points,one,a method for determining the number of anchor boxes based on hierarchical clustering algorithm is proposed.The other,for the characteristics of large license plate targets,the multi-scale feature fusion of the network is optimized.The optimized PLATE-YOLO network is more suitable for license plate detection and improves the detection speed.The experiment results show that the PLATE-YOLO network detects the speed of the license plate by 5 FPS compared with the YOLOv3.In the smog environment,the PLATE-YOLO network after dehazing pretreatment is more accurate than the unlicensed license plate detection method 9.2%.