轻量化YOLOv7-tiny的水下压印字符识别
Lightweight YOLOv7-tiny underwater embossed character recognition
李卓润 1李波 1邱鹏程 1刘洪1
作者信息
- 1. 中国地质大学(武汉)机械与电子信息学院 武汉 430074
- 折叠
摘要
自动化水下字符识别技术能通过编号更高效地定位追踪水下设备,是管理和维护水下设备的关键.针对该任务目标区别较小和水下场景中干扰等问题,并考虑其检测速度需求,基于YOLOv7-tiny模型,提出一种轻量化的改进模型.首先采用 MobileNetV3作为新的特征提取网络对整体框架进行轻量化处理;然后引入PConv至ELAN模块中,减少Neck层的计算量;最后将置换注意力机制应用至Head层,提升了模型对字符定位的表达能力.实验结果表明,改进后的模型相较于原模型的平均精度均值(mAP)提高了2.4%,参数量和计算量分别减少30.0%和38.5%,检测速度提升30.8%.改进后的模型在水下字符识别任务中具有更高的效率和精度,为推进并实现水下自动化识别编号设备的部署提供了可行性.
Abstract
Automated underwater character recognition technology can more efficiently locate and track underwater equipment through numbers,which is the key to managing and maintaining underwater equipment.In view of the problems such as the small target difference of the task and interference in the underwater scene,and considering its detection speed requirements,this article proposes a lightweight improved model based on the YOLOv7-tiny model.First,MobileNetV3 is used as a new feature extraction network to lightweight the overall framework.Then PConv is introduced into the ELAN module to reduce the calculation amount of the Neck layer.Finally,the displacement attention mechanism is applied to the Head layer to improve the model's ability to position characters.expression ability.Experimental results show that compared with the original model,the mAP of the improved model is increased by 2.4%,the amount of parameters and calculations are reduced by 30.0%and 38.5%respectively,and the detection speed is increased by 30.8%.The improved model has higher efficiency and accuracy in underwater character recognition tasks,providing feasibility for promoting and realizing the deployment of underwater automated identification and numbering equipment.
关键词
水下字符识别/YOLOv7-tiny/轻量化/PConv/置换注意力Key words
underwater character recognition/YOLOv7-tiny/lightweight/PConv/shuffle attention引用本文复制引用
基金项目
国家重点研发计划(2023YFC2813104)
国家重点研发计划(2023YFC3007004)
出版年
2024
NETL
NSTL
万方数据