Research On Recognition Method Of Sound Level Meter Reading Based On Deep Learning

In order to deal with the noise pollution hazards generated by transportation,construction,industrial production and other industries,our country has continuously strengthened financial investment,improved relevant policies,and actively carried out noise pollution prevention and control work.As an important noise measuring instrument in noise pollution prevention and control,the sound level meter often needs to be sent to the measurement department for calibration to ensure the accuracy of its measurement.The staff needs to manually read and record the sound level meter readings during the calibration process,and the traditional sound level meter reading recording method will consume a lot of manpower and time,and it is inevitable that there will be problems such as wrong records and omissions.In view of the above problems,this paper studies the reading detection and recognition method for sound level meter images based on deep learning,and designs and implements a sound level meter reading recognition system,the main work contents are:1.A light-weight sound level meter reading detection method based on DBNet(Differentiable Binarization Network)network is proposed.In this method,the lightweight network Shuffle Net V2 is used as the backbone network,which greatly reduces the amount of parameters and calculations of the model.At the same time,the efficient channel attention ECA(Efficient Channel Attention Module)module is integrated into the feature pyramid network,and the channel information of the feature map is enhanced through the weighting strategy of the ECA module,which improves the extraction ability of the network for important features,effectively improves the accuracy of the model,and only increases a small amount of calculation.The experimental results on the sound level meter reading test dataset show that the optimized model parameters are reduced to 15.4%,the calculation amount is reduced to 67.4%,and the F1 value reaches 0.982,which is an increase of0.9% compared with the original method,and the detection time of each image is about 0.372 s.2.Improved a sound level meter reading recognition method based on CRNN(Convolutional Recurrent Neural Network)network.Firstly,a batch normalization layer is added to the feature extraction network to improve the stability during network training.Then,the residual block is introduced to replace the original convolutional block,which deepens the depth of the network and improves the extraction ability of the network for complex features.Finally,Dropout is applied to the feature extraction network to suppress the overfitting problem.In addition,the model is pre-trained by fine-tuning technology,which effectively improves the model accuracy and network convergence speed.The experimental results on the sound level meter reading recognition dataset show that the accuracy of the improved method reaches 99.7%,which is 2.4% higher than the original method,and the recognition time of a single reading image is about 0.025 s.3.A sound level meter reading recognition system is designed and implemented.The sound level meter reading recognition system is a windowed software built based on Py Qt5,Python and Open CV,and the main functional modules include image automatic acquisition module and reading image recognition module,which realizes the function of automatic acquisition,recognition and recording of sound level meter images.The test results show that the system can assist the staff in the daily calibration of the sound level meter.