Corrosion Detection at Transmission Accessories Using Combination of Object Detection, Image Classification and Background Removal

Abstract

Inspection of electrical transmission accessories is an important aspect of maintaining the reliability of electricity supply. But due to the large number of components that need to be inspected and cost to maintain it, lately there has been a trend to inspecting it by autonomous using AI or Deep Learning. And one of state of the art in object detection model is YOLO and its variant from Ultralystic. Using it in object detection and image classification task using new dataset and exclusive form Indonesia environment that researcher never has been done, that is Clevis, Dead end, Shackle, Tension clamp, Hole and Bolt. Datasets are also provided in several types to target the highest score. Also use another model like YOLOX and RT-DETR for object detection, and VGG-19, DenseNet-201, etc. to image classification. And the result is that smaller number of class and using larger object can improve metric score result overall, and it get by model YOLOv9e, it can reach 0,972 in mAP@0,5. And the removal of background will lead to poor metric score, even it cannot reach 0,20 poin in mAP@0,5 except using YOLOX-X that reach 0,661. Combination between object detection and image classification seems good at training and testing part, but when it is used with original data the result decreases, from about 0,94~0,97 to 0,82~0,84 precision. This is likely due to IoU limitations when extracting it from object detection which causes lower quality dataset that deliver to image classification process.