Research On Traffic Mark Recognition Based On Distributed Deep Learning

Computer vision is an interdisciplinary subject that studies how to understand the highlevel semantics of digital images or videos,required the ability to see in the human brain,mainly in the visual cortex.As a basic task of computer vision,object detection plays an important role in understanding pictures.Target detection refers to the process of automatically locating certain or certain objects in a given image according to the requirements of people.In order to achieve this function,we often use some model frameworks of deep learning in both scientific research and engineering.The research of traffic mark recognition based on deep learning proposed in this paper is to use the method of computer vision field to identify and segment the traffic mark image.The whole process includes five steps: image preprocessing,image edge extraction,image positioning,image segmentation and image recognition.The source data used in this paper is point cloud data,so we first convert the point cloud image into grayscale image.In the preprocessing stage of the image,image processing technology and image enhancement algorithm are used,such as airspace filtering,frequency filtering,histogram equalization,etc.,mainly to facilitate the rapid and accurate extraction of the line edge from the road background image.After determining the edge of the line,the computer can roughly determine the effective position of the line and then carry out image segmentation and recognition operations.Different from traditional image recognition,the recognition method proposed in this paper is based on multi-node distributed operation,which can greatly improve the speed and accuracy of image recognition.This paper adopts Tensor Flow on Spark,a distributed deep learning framework opensource by yahoo.Tensor Flow on Spark brings an extensible deep learning approach to big data computing clusters by combining the deep learning framework Tensor Flow with the big data framework Apache Spark and Apache Hadoop,It also enables Tensor Flow to implement deep learning on GPU or CPU server clusters.Tensor Flow on Spark is a mature distributed deep learning framework,which has been widely used in image recognition,NLP and speech recognition.Its algorithm has been highly optimized,which can give full play to the computing advantages of the classical deep learning framework Tensor Flow,fast computation speed,can carry out efficient and fast deep learning ability on the existing computer cluster.In the experiment,we first constructed Spark distributed framework,and then directly used GPU to do distributed model training.Tensor Flow on Spark supports all current popular Tensor Flow applications,and any Tensor Flow application can be easily ported to Tensor Flow on Spark.In addition,Tensor Flow on Spark supports model parallelism and data parallelism,making it easier to implement asynchronous training and prediction.In the experiment of this paper,the converted gray scale map is first divided into ten categories for marking,which constitute a training sample set.Then the training set is input into the established distributed deep learning framework to train the discriminant model.Finally,pictures were randomly selected to constitute the test set to verify the recognition effect of the discriminant model,and the recognition results were counted according to the three criteria of accurate discrimination,misjudgment and omission.The results show that the method proposed in this paper has high accuracy in the identification of traffic marks,and is practical in engineering and scientific research.