基于改进DBNet-RNN的声级计读数识别方法  

Sound Level Meter Reading Recognition Method Based on Improved DBNet-CRNN

作　　者：汪嘉 祝海江[1] 王寅初 何龙标[2] 杨平[2] 牛锋[2] WANG Jia;ZHU Haijiang;WANG Yinchu;HE Longbiao;YANG Ping;NIU Feng(College of Information Science and Technology,Beijing University of Chemical Technology,Beijing 100029,China;National Institute of Metrology,Beijing 100029,China)

机构地区：[1]北京化工大学信息科学与技术学院,北京100029 [2]中国计量科学研究院,北京100029

出　　处：《计量学报》2024年第8期1191-1199,共9页Acta Metrologica Sinica

基　　金：国家重点研发计划(2021YFF0600202)。

摘　　要：为提高声级计校准工作效率,提出了一种基于深度学习神经网络的声级计图像读数检测与识别方法。读数检测模型以DBNet为基础模型,将ShuffleNetV2作为主干网络,显著降低模型参数量;为提高读数区域检测精度,引入高效通道注意力ECA模块,提高网络对于通道特征的提取能力,优化后的模型在保持精度的同时参数量缩减为原来的15.4%,计算量缩减为原来的67.4%。读数识别模型以CRNN为基础模型,先加入批量规范化层,提高网络训练时的稳定性;然后,引入残差块替换原有的卷积块,提高了网络对于复杂特征的提取能力;将Dropout应用于网络中,提高网络的泛化能力;此外,在合成读数数据集上对读数识别模型进行预训练,有效增加了模型准确率。改进后的方法准确率达到了99.7%,相较原方法提高了2.4%。实验结果表明,该方法对声级计图像中存在的字体多样、光照不均、模糊等影响因素具有较强的鲁棒能力,对声级计图像中的读数具有较高的识别精度。In order to improve the work efficiency of the sound level meter calibration work,a reading detection and recognition method based on deep learning neural network for the image of the sound level meter is proposed.The reading detection model is based on DBNet and uses ShuffleNetV2 as the backbone network,significantly reducing the number of model parameters.To improve the accuracy of reading area detection,an efficient channel attention ECA module is introduced to enhance the network's ability to extract channel features.The optimized model reduces the number of parameters to 15.4%and the calculation amount to 67.4%while maintaining accuracy.The reading recognition model is based on the CRNN model,which first adds a batch normalization layer to improve the stability of network training.Then,residual blocks are introduced to replace the original convolutional blocks,improving the network's ability to extract complex features.Applying Dropout to the network to improve its generalization ability.In addition,pre-training the reading recognition model on the synthesized reading dataset effectively increases the accuracy of the model.The improved method achieves an accuracy of 99.7%,which is 2.4%higher than the original method.The experimental results show that this method has strong robustness against factors such as diverse fonts,uneven lighting,and blurring in sound level meter images,and has high recognition accuracy for readings in sound level meter images.

关 键 词：声学计量 图像识别 声级计 DBNet CRNN 注意力模块 

分 类 号：TB95[一般工业技术—计量学]