大角度透视变形集装箱图像箱号识别方法

Container Code Recognition from Images with Large Perspective Deformation

摘要:

针对存在大角度透视变形的集装箱图像，提出一种新的集装箱箱号识别方法.首先对图像进行透视变换校正，然后利用深度卷积神经网络模型定位并识别出集装箱图像中的26个大写英文字母和10个阿拉伯数字，最后利用集装箱箱号的先验知识，通过级联决策规则从候选字符集中识别出集装箱箱号.此方法应用于重庆港集装箱1 035张实景图像，箱号识别精度达97%，基于NVIDIA GeForce GTX1080图形处理器加速的箱号识别速度为每秒2~5帧.

Abstract:

A novel method is proposed in this paper to recognize the container code from the images with large perspective deformation. First, the images are rectified by perspective transformation. Then, 26 capitalized English characters and 10 Arabic numerals are located and recognized based on the deep convolution neural network model. Finally, container codes are recognized from the candidate character set by cascade decision rules based on the priori knowledge of container code. The proposed method is verified by 1035 container images taken in Chongqing Port. The result shows that the accuracy of container code recognition reaches 97%, and the speed based on NVIDIA GeForce GTX1080 GPU is 2 to 5 frames/sec.