作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• •    

基于改进PPYOLOE-R的信息码矫正研究

  • 出版日期:2023-10-30 发布日期:2023-10-30

Research on information code correction based on improved PPYOLOE-R

  • Online:2023-10-30 Published:2023-10-30

摘要: 信息码识别技术推动着社会的进步,使人们的生活更便捷。由于所处拍照环境影响,信息码识别效果有待提高,而且信息码角度倾斜也会影响解码正确率,本文以基于信息码的电力互感器误差试验接线判断为背景,在AI边缘计算设备上进行改进PPYOLOE-R的信息码矫正研究。首先以PPYOLOE-R检测算法为基础,融合轻量级网络ESNet,在提升精度同时降低模型参数量;其次,引入动态卷积进一步加强特征提取,减少模型因下采样中的信息丢失,加强模型通道特征提取能力;最后为满足AI边缘设备上实时性要求,采用模型融合技术将推理模型进行融合,保证模型精度不变的情况下,提升模型检测速度。为丰富数据集,采用两步旋转数据增强和Mosaic + Mixup数据增强方法,充分利用数据集中已有信息,提高模型学习能力。实验表明,改进后算法精度达到89.46%,较原模型提升了1.95%,检测照片速度从154ms每张提升至50ms每张。相较其他算法,所改进的算法有着体积小和速度快优势,通过算法矫正后的信息码,可显著提高解码效率和正确率。

Abstract: Information code recognition technology promotes the progress of society and makes people's life more convenient. Due to the impact of the photography environment, the information code recognition effect needs to be improved, and the information code Angle tilt will also affect the decoding accuracy. Based on the information code-based power transformer error test wiring judgment as the background, this paper studies the improvement of PPYOLOE-R information code correction on AI edge computing equipment. First, based on PPYOLOE-R detection algorithm, lightweight network ESNet is integrated to improve accuracy and reduce the number of model parameters. Secondly, dynamic convolution is introduced to further enhance feature extraction, reduce the information loss in the model due to subsampling, and enhance the feature extraction capability of the model channel. Finally, in order to meet the real-time requirements on AI edge devices, the model fusion technology is used to fuse the inference model, so as to improve the model detection speed without changing the accuracy of the model. In order to enrich the data set, two-step rotation data enhancement and Mosaic + Mixup data enhancement methods were used to make full use of the existing information in the data set and improve the learning ability of the model. Experiments show that the accuracy of the improved algorithm reaches 89.46%, which is 1.95% higher than that of the original model, and the detection speed is increased from 154ms per photo to 50ms per photo. Compared with other algorithms, the improved algorithm has the advantages of small size and fast speed, and the decoding efficiency and accuracy can be significantly improved through the corrected information code.