基于级联多任务深度学习的卡口识别引擎研究

Study on Bayonet Recognition Engine Based on Cascade Multitask Deep Learning

针对在将卡口非结构化视频图像数据转化为智能结构化信息的过程中存在环境的复杂性、需求的多样性、任务的关联性和识别的实时性等问题,提出了一种级联多任务深度学习网络的卡口识别引擎方法,其通过充分利用分割、检测、识别等任务之间的相互联系实现了高精度的、高效的、同步实时的卡口车辆多种基本信息的识别(车型、品牌、车系、车身颜色以及车牌等识别任务)。首先,利用深度卷积神经网络自动完成车型的深度特征学习,在特征图上进行逻辑回归,从卡口道路复杂背景中提取出感兴趣区域(包括多车辆对象);然后,利用多任务深度学习网络对提取出来的车辆对象实现多层次的多任务识别。实验结果表明,提出的方法在识别精度和效率上都明显优于传统计算机视觉方法和现有的基于深度学习的识别引擎技术,该方法对车型、品牌、车系及车牌的识别与检测精度均达到98%以上,检测效率提升了1.6倍。

Aiming at the complexity of environment,the diversity of requirements,the relevance of tasks and the real-time of identification in the process of converting the unstructured video data of bayonet into the intelligent structured information,this paper proposed a method of bayonet recognition engine based on cascade multitask deep learning.This method makes full use of the relationship between segmentation and detection recognition tasks to achieve high-precision,efficient,synchronous and real-time recognition of a variety of basic information of bayonet vehicles(motorcycle types,brands,series,colors and license plates etc.).First,the deep convolutional neural network is used to automatically extract the depth feature and the logical regression is performed on the feature map to extract the interested region from the complex background(including multi-vehicle object).And then the multitask deep learning network is used to achieve multilevel multitask recognition for the extracted vehicle objects.Experimental results show that the proposed method is superior to the traditional computer vision method and the existing recognition engine technology based on deep learning in terms of recognition accuracy and efficiency,and the accuracy of recognizing and detecting the motorcycle types,brands,series and license plates is more than 99%respectively,and the detection efficiency is increased by 1.6 times.