Abstract:
Water agencies often record water levels daily using a telemeter or staff gauge. Telemeters can automatically measure water levels, but they are expensive. Staff gauges are less expensive, but they cannot automatically record data. Many agencies now stream staff-gauge images from closed-circuit television to the public.
This thesis presents a method for applying the Convolutional Neural Network technique to detect water levels from staff gauge images streamed from closed-circuit television. Images of staff gauges from the streaming websites are recorded every 60 minutes. Then, those images are used to train a model to detect all numbers, the E symbols in the staff gauge and to detect the staff gauge pole. If there are multiple poles, the proposed method selects and reads the value of the water level on the smallest pole size.
The experimental results showed that the proposed method had an average accuracy of 83.52%, which was significantly higher than the Optical Character Recognition method. The Optical Character Recognition method cannot detect any numbers, staff gauges, or the E symbols. The Convolutional Neural Network with Hue, Saturation and Value method is able to detect the numbers close to the water level with an average of only 27.45%. Furthermore, this proposed method is able to detect up to 3 or more water level gauges, while the OCR and CNN+HSV methods could detect only a single gauge.