Research On Reading Recognition Of Wheel-type Water Meter Based On Deep Learning

In order to reduce the work intensity and improve the efficiency of reading on the word wheel type water meter,people use computer vision to automatically identify the readings in the image.In particular,the rapid development of deep learning has greatly promoted the application of automatic recognition of water meters in various scenarios.However,due to the long-tailed distribution of the class,small objects and incomplete numbers,which affect the performance of the automatic recognition of water meter based on deep learning.we propose a Mix suppression data augmentation algorithm and a block attention module combined with deep learning technology to accurately detect water meter readings in long-tailed distribution datasets.First,a mix suppression data augmentation method based on the information deletion and resampling is proposed the method is divided into two parts.The first part expands the number of incomplete digital samples and improves the diversity of the characteristics of the incomplete digital samples.In the second part,the training times of the head class label and the tail class label are gradually balanced during the training process,so as to improve the long-tailed distribution phenomenon of the water meter dataset and the detection performance of small objects.Second,in order to improve the detection performance of convolutional neural networks for object concentrated regions,a block space attention module is proposed to connect the output of the backbone network.This module can enhance the information mining ability of the label concentrated area,and the calculation cost is extremely low.Third,based on the Android platform,an intelligent meter reading application was developed,and the water meter detection model proposed in this paper was implemented and deployed on the mobile terminal.The experiments show that our methods successfully improve the performance of neural networks for the water meter digit reading recognition,and they could also be conveniently embedded into other mainstream general object detection models.On the water meter dataset established in this paper,it achieves 97.2% m AP and 94.1% F1.