YOLO-Based Valve Type Recognition and Localization
Vahid Behtaji Siahkal Mahalleh, Tareq Aziz AL-Qutami, Iskandar Al-Thani Mahmood · 2019 IEEE 6th International Conference on Industrial Engineering and Applications (ICIEA) · 2019
In this paper, we present a YOLO model to recognize valve types and localize their positions on an experimental test bench unit in real-time frames taken from a continuous stream of video data, which are captured by a camera in a hydrogen sulfide (H2S) laboratory. The model required input frames after a particular period. In additional, the model was able to determine the type of valve labels based on a single frame. We showed that the pre-processing of training images plays an important role in the performance of YOLO model. We also included the concept of anchor and intersection over union (IOU) for good accuracy. Valve type results were demonstrated over a specific time of the video stream. The achieved average loss error value of the model was less than 6 percentage and the IOU between the ground-truth bounding box and predicted bounding box for all regions in the final iterations was almost 95 percentage.