A Text Recognition System For Reading Meters
T his stud y gives a solutio n for auto matic reading of meters b y taking digital images o f the meter s and using machine lear ning techniq ues to auto matically read the meter s. A dataset of positive examp les, negative examples and test digital images is collected and used to train a classifier using the Adaboo st learning algorithm. Positive examples are digital images o f meters collected rando mly fro m manually read meters and negative examples are any digital images without meter s taken rando mly using a digital camer a, collected fro m past images or har vested o n the inter net. T he test data is a subset of the positive and negative examples, and is not used in the algorithm training process. T he p ositive and test images are manually labelled with the regio ns that contain meters on those images and a descr iptio n file is created with the coordinates o f meters in the images. T he stud y uses OpenCV libraries for the classifier training and validatio n. T he results o f the the best perfor mance test o n the test data fro m the Adaboo st training sessions is 75.4% recall with 19.4% false positive rate. T his classifier is then used in a prototype that read s meters fro m digital images. Once the meter region is detected on an image, temp late matching techniq ues are used to locate the meter reading. T he meter reading region is then binarized and presented to an OCR engine for text reading. It is this reading that can then be transmitted to utility providers for custo mer billing. Various integrity checks are proposed to ensure the transmitted data is accurate to avoid losses to utility provider s and custo mer s. T he stud y presents this prototype to demo strate the ability to use current technolo gies to solve commo n problems withi n society.