Research on Low-Resolution Pattern Coding Recognition Method Based on Hu-DBN

基于Hu-DBN的低分辨图案编码识别方法研究

¹ School of Economics and Management, Hubei Engineering University, Xiaogan 432100, China
² School of Automation, Northwestern Polytechnical University, Xi'an 710072, China
³ Shaanxi Xueqian Normal University, Xi'an 710072, China

Received: 20 November 2019

Abstract

The feature image code represented by the two-dimensional code is the key reference for global positioning in the visual navigation of mobile robots. Although reducing the acquired low-resolution image helps to reduce the real-time performance of the algorithm, the acquired feature image is more susceptible to motion blur-based interference and affects the accuracy of recognition, which causes the positioning failure of the whole multi-intelligence, in which the body control system is invalid. In this paper, an optimized low-resolution feature image code recognition method is proposed. In the preprocessing part, the characteristic image is converted into the characteristic signal matrix of Hu invariant moments, and then the characteristic image is added to the characteristic signal matrix as a characteristic component, and then the Hu-DBN neural network signal classifier is used to construct the signal matrix so as to achieve accurate recognition of low-resolution custom image signature images under high motion tolerance conditions. It not only avoids the problem of classical pattern recognition relying on model experience and poor adaptability of the scene, but also avoids the problem of high computational complexity and recognition efficiency of directly deep learning methods such as YOLO. The deployment of the mobile robot instance deployment test shows that the average recognition rate is of 96.3% at a resolution of 640×480@Pixs and motion speed of 0.5 m/s, which proves the effectiveness of the present method.

摘要

图案编码是移动机器人视觉导航中全局定位的关键参照物。通过降低图像编码的影像尺寸和质量有助于减少运算量，以提升算法实时性，但对应特征图像更易受以运动模糊为主的干扰而影响识别的准确性，使得定位失败进而造成整个多智能体控制系统失效。提出了一种优化的低分辨率特征图像码识别方法，在预处理部分将特征图像转换为其Hu不变矩的特征信号矩阵，再将特征图像作为一个特征分量补充到该特征信号矩阵中，通过构建的Hu-DBN神经网络信号分类器对信号矩阵进行学习，从而实现低分辨率自定义图像特征码影像在较高运动容差条件下的准确识别。既避免了经典模式识别依赖模型经验、场景适应性差的问题，同时也避免了直接使用如YOLO等深度学习方法带来的高运算量和识别效率问题。通过移动机器人实例部署展开测试，通过搭载分辨率为640×480图像传感器移动机器人实例展开测试，在0.5 m/s运动速度时平均识别率为96.3%，证明了所提方法的有效性。