Evaluation of Recognition of Water-meter Digits with Application Programs, APIs, and Machine Learning Algorithms
Kwanchai Eurviriyanukul, Kriatsanga Phiewluang, Sirisak Yawichai, Sirilak Chaichana · 2020
Image recognition of digits of water meters is useful and timesaving. Many exist programs/APIs available for image recognition but not specially for this task. Therefore, this paper investigated and compared correctness of those 5 programs/APIs, i.e. Anyline, Line, Google Vision, Microsoft Azure Computer Vision, and Naver (Clovar) OCR with 32 water-meter images with 3 different perspectives (straight, upside-down, right-rotation). In addition, those images were manually cropped into single-digit images. Then, they were tested with AutoML, Tesseract and some machine learning algorithms (KNN, SVM and CNNs). trained with 1,500 pictures (28*28 pixels) of a single digit of water meter images. The experimental results showed that Anyline got 96.9% of accuracy for the straight images. Whereas, GoogLeNet got 78.9% and 74.2% for right-rotation and upside-down images, respectively. On the other hand google vision, Anyline and Line got 34.4%, 0% and 18.8% accuracy for straight, right-rotation and upside-down images, respectively. In addition, dataset of EMNIST and Google Street View were used for training of CNN. However, about 12% and 40% correctness were achieved for those dataset, respectively.