Research And Application Of Character Wheel Water Meter Reading Recognition Based On Deep Learning

Water meter images taken in natural scenes have different lighting conditions,stains and shooting angles,which have a great impact on the detection and recognition of reading areas.In addition,the edge characteristics of the box in the reading area of different types of water meters are greatly different,and the number of characters in the reading is also different.These factors further increase the difficulty of reading recognition.Although the current reading detection and recognition methods of the character wheel water meter have achieved some research results,there are still some shortcomings in solving the above problems,such as insufficient generalization ability of the method,easy to be interfered by environmental factors and large recognition model,which can’t meet the requirements of reading detection and recognition of the water meter in natural scenes.To solve the above problems,this article studies the reading detection and recognition algorithm of character wheel water meter based on deep learning,and applies it to recognize water meter images in different environments,and achieves good expected results.The main work is as follows:1.A preprocessing method of large resolution water meter image based on regional interpolation method is adopted,so that the feature information of the original image can be retained after scaling the large resolution water meter image without resulting in image distortion.At the same time,the mean value of the scaled water meter image is used to adjust the brightness and contrast to enhance the water meter image.2.The water meter disc detection is regarded as target detection,and the Yolov4 target detection algorithm is used to detect and locate the water meter disc image.The training method of transfer learning is adopted to solve the overfitting problem caused by the small number of training samples,accelerate the convergence of the network and improve the performance of the model.By adjusting the output frame of the disc detection and using the perspective transformation algorithm to transform the water meter disc images taken from different angles,the image feature information in the reading area of the segmented water meter disc images can remain unchanged after scaling.3.An adaptive histogram equalization algorithm with limited contrast and bilateral filtering algorithm are introduced to enhance the image of the water disc.At the same time,a reading region detection algorithm based on full convolution is designed to complete the detection of reading regions of water meters with different rotation directions without the assistance of traditional methods.The experimental results show that the trained network model has a better detection effect on the reading region.4.According to the detection results of reading area and the size information of water meter disc image,a water meter reading area correction and segmentation algorithm is designed to accurately segment the image of reading area.For the image of reading region,a dynamic character segmentation algorithm based on the combination of projection method and sliding window method is proposed to achieve character segmentation.Finally,a lightweight reading recognition network is designed to recognize water meter reading.The experimental results show that it requires less model parameters and training time,and at the same time maintains a high recognition accuracy.5.Design the water meter reading detection and identification system based on the reading detection and identification method in this article.Firstly,Tkinter is used to design a visualization platform to test the feasibility and performance of the proposed method.Then according to the test results to design the water meter reading detection and recognition system.The test results show that the method in this article can complete the identification of water meter readings in about 0.3 seconds,and the accuracy of different types of water meter readings in complex environments is 95.176%.With the exception of occluded or missing readings,all other water meter readings can be identified to achieve the desired effect.