Research And Implement Of Digital Identification Of Water Meter

With the rapid development of computer technology, a lot of traditional work methods have become time consuming and low efficiency. Thus semi-automatic or automatic methods gradually replace the artificial methods. In traditional ways, it needs people to record the water consumption of the users. This method takes a long time,requires a lot of manpower, and it is easy to make mistakes in the process of meter reading.According to these disadvantages of artificial means, it is very necessary to adopt automatic meter reading technology, and the most important is the accurate and rapid identification of the digital water meter.There exist some problems such as aging, corrosion and spot, stains and others, and thus the water meter image will be more blurred. For this type of image, the first thing to do is the preprocessing of the image, which including image filter, image tilt correction,image binarization and the character segmentation, etc.. After the preprocessing of the image,we need to remove the useless information, and separate out the digital character from the water meter image.Hog(histogram of oriented gradient) feature is a descriptor which can be used in computer vision and image processing for object detection. And it is composed by means of computing object statistics of a local region of the gradient histogram, which is a very good descriptor of the object edge features. SVM(Vector Machine Support) is one of the best classifiers, which is commonly used to classify objects. The combination of HOG and SVM can effectively realize the recognition of digital characters. Firstly tidied the ten digit samples from zero to nine, and then use them to train the classifiers. After the water meter image preprocessing, we get a set of the separated digital of the water meter image, and normalize each digital, then extract hog feature of each digital. After all the things have be done, we use the classifier which has been trained to predict the digital, and then we combine the results and take it as the final forecasting result. For the prediction that is inaccurate, we train the digital again. In theory, as long as the number of samples is large enough, the recognition rate can reach 100%.Finally, a high recognition rate is obtained by the test of a large number of water meter images, which verifies the feasibility and effectiveness of the system.