摘要
针对中国象棋机器人娱乐化智能设备,设计了其关键的视觉系统,实现棋子定位与棋子识别。棋子识别是关键的一步,针对传统识别方法的局限性,以及棋子旋转特性带来的难识别问题,提出一种基于深度学习的字符识别方法。以分组卷积方式改进Inception-V3网络结构,融合可变形卷积层,增强字符形变特征提取能力。实验显示,模型的识别准确率达到99.99%,比传统方法识别效果更好,说明了该方法在棋子识别任务上的有效性。
The key vision system of Chinese chess robot entertainment intelligent equipment is designed to realize chessman positioning and chessman recognition. Chessman recognition is a key step. Aiming at the limitation of traditional recognition method and the difficulty of chessman rotation, a character recognition method based on deep learning is proposed. The Inception-V3 network structure is improved by grouping convolution, and the deformable convolutional layer is fused to enhance character deformation feature extraction capability. The experiment shows that the recognition accuracy of the model is 99.99%, which is better than the traditional method.
作者
孔凡国
李智宗
刘庆
黄伟
KONG Fanguo;LI Zhizong;LIU Qing;HUANG Wei(Division of Intelligent Manufacturing,Wuyi University,Jiangmen 529020,China;Yuanmeng Precision Technology(Shenzhen)Institute,Shenzhen 518110,China)
出处
《机械工程师》
2020年第5期1-3,共3页
Mechanical Engineer
基金
广东省引进创新创业团队项目(201301G0105337290)
深圳市海外高层次人才创新创业专项资金(20130412193329296)。
关键词
象棋机器人
棋子识别
可变形卷积
深度学习
chess robot
chess recognition
deformable convolution
deep learning
