基于深度学习的车检图像多目标检测与识别

Multi-target Detection and Recognition for Vehicle Inspection Images Based on Deep Learning

为了实现快速和自动的车辆外观检测,提出一种基于深度学习的车检图像多目标检测与识别方法。首先,采用轻量级神经网络YOLOv3实现车检图像中车头、轮胎、车牌及三角形标志的检测与识别;其次,采用多任务级联卷积神经网络实现车牌4个关键点定位;再次,利用车牌4个关键点坐标,结合目标车牌图像高宽先验,通过透视变换对车牌进行校正;最后,设计卷积神经网络实现车牌底色分类,同时设计卷积循环神经网络,实现车牌字符识别。实验结果表明,在816×612的车检图像上,该方法中端到端的多目标检测与识别的平均精度达98.03%;为便于在车检场景下应用该模型,利用阿里巴巴推理引擎将模型部署到CPU端,使多目标检测与识别的平均速度达10帧/s,从而满足车检的应用需求。

A multi-target detection and recognition method of vehicle inspection images based on deep learning is proposed for faster and more automatic vehicle inspection. Firstly,a lightweight yolov3 network is used to detect and recognize vehicle head, tires, license plate and triangle marks in a vehicle inspection image;secondly, a multi-task cascade convolution neural network is used to locate the four key points of the license plate;thirdly, according to the four key point coordinates and the size prior of the target license plate, the license plate image is corrected by perspective transformation;finally, a convolutional neural network is designed to classify the background color of the license plate. Thus, a convolutional recurrent neural network is realized for license plate character recognition. Experimental results show that the average end-to-end multi-target detection and recognition accuracy of this method is 98.03% on an 816 × 612 car inspection image. To facilitate the application of the deep learning model in vehicle inspection scenes, the model is deployed to a CPU using Alibaba reasoning engine, and the average speed of multi-target detection and recognition reaches 10 frames per second, which meets the application requirements of vehicle inspection.