AUTOMATIC CALIBRATION USING YOLOV5 AND DEEP HOMOGRAPHY - PERSPECTIVE CORRECTION IN WATER LEVEL MEASUREMENT SYSTEM

• Main Author: Luscovius Purba, Poppy
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/69982
• Institution: Institut Teknologi Bandung

This study discusses the application of several neural networks with minimal computational requirements to be able to calibrate perturbated images based on template images. The need for this research is to reduce the cost of observing water level which generally uses conventional sensors which require higher maintenance costs compared to the use of Artificial Intelligence (AI) and CCTV cameras as sensors. To overcome the large internet needs, image processing is carried out on site using a neural network which can reduce computational needs. The initial stage is to detect symbols that have been designed with YOLOv5 and validate the YOLOv5 results to ensure the sequence of symbols matches the template image. YOLOv5 results that have been validated will be used as input data for Squeeze-HomographyNet which has been developed and trained to perform perspective correction on perturbated measuring board images. The results of the calibrated image will be estimated as a cut-out of the measuring board in the main image to reduce computational requirements in the water surface line detection phase. Furthermore, in the measuring board image the water surface line will be detected and the distance measured with the closest validated symbol and a comparison scale is used between the actual distance (cm) and the image distance (pixels) to measure the entire water level in the observation image. The results of the system development find that in the best case the system can produce 90.77% accuracy for the water level prediction. The resulting neural network model is quite small, around 41.165 Kb for YOLOv5 and 10.987 Kb for Squeeze-HomographyNet.