计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (20): 165-172.DOI: 10.3778/j.issn.1002-8331.1908-0472

• 模式识别与人工智能 • 上一篇    下一篇

多尺度特征融合的二维编码图像识别

续晓丽,刘从峰,范广宇   

  1. 1.中国科学院 上海技术物理研究所 红外成像材料与器件重点实验室,上海 200083
    2.中国科学院 上海技术物理研究所 传感技术国家重点实验室,上海 200083
    3.中国科学院大学,北京 100049
    4.常州光电技术研究所,江苏 常州 213100
  • 出版日期:2020-10-15 发布日期:2020-10-13

Recognition of Two-Dimensional Coded Image Based on Multi-scale Feature Fusion

XU Xiaoli, LIU Congfeng, FAN Guangyu   

  1. 1.Key Laboratory of Infrared Imaging Materials and Detectors, Shanghai Institute of Technical Physics, Chinese Academy of Sciences, Shanghai 200083, China
    2.State Key Laboratories of Transducer Technology, Shanghai Institute of Technical Physics, Chinese Academy of Sciences, Shanghai 200083, China
    3.University of Chinese Academy of Sciences, Beijing 100049, China
    4.Changzhou Institute of Optoelectric Technology, Changzhou, Jiangsu 213100, China
  • Online:2020-10-15 Published:2020-10-13

摘要:

为实现二维编码同时适合机器和人的识读,提出了一种二维编码图像结合的新型二维图像编码技术,并针对该二维编码图像无定位标识的特点,采用目标识别的算法进行识别,同时改进了SSD(Single Shot Multibox Detector)算法使其适用于此类目标识别。该算法针对小目标的识别特点,去掉了原有的空洞卷积层以及选择了浅层的图层进行特征融合,达到了准确识别二维编码图像的效果。此外搭建了基于GTX1080型号的GPU、Ubantu Linux和TensorFlow的深度学习平台,提出了一种网络学习和验证方法,该方法通过电脑生成相应的数据集来训练网络,并采用合成的数据集和真实图片分别对网络进行验证和测试。测试结果表明,该方法训练的网络对具有良好拍照条件的真实图片的编码识别准确率达到100%。

关键词: 二维编码, 目标识别, SSD算法, Tensorflow

Abstract:

In order to realize two-dimensional coding and fit for machine and human reading, a new two-dimensional image coding technology combining two-dimensional coded images is proposed. According to the characteristics of the two-dimensional coded image without location identification, the target recognition algorithm is adopted to recognize the two-dimensional coding image, while improving the SSD(Single Shot Multibox Detector)algorithm to make it suitable for such target recognition. According to the recognition characteristics of small targets, the algorithm removes the original empty convolution layer and selects the shallow layer for feature fusion, and achieves the effect of accurate recognition of two-dimensional coded image. In addition, the depth learning platform based on the GTX1080 model is built, and a network learning and verification method is proposed, which trains network by generating corresponding data sets by computer, and verifies network separately by using the synthesized data set and real pictures. The test results show that the trained network by the proposed method has a coding recognition accuracy of 100% for real pictures with good photographing conditions.

Key words: two-dimensional coded, target recognition, Single Shot Multibox Detector(SSD), Tensorflow