An Automatic Respirator Oil Level Reading Method Based on Instance Segmentation

摘要

The promotion of intelligent video surveillance technology has promoted the demand for intelligent image recognition algorithm in the power industry. Power transformer is a very important equipment in substation. In order to realize the automatic reading of the oil level for the transformer respirator, a reading method based on instance segmentation is proposed. Firstly, a Mask Intersection-over-Union (MaskloU) regression module is added to the standard Mask R-CNN model in order to refine the mask confidence and improve the segmentation accuracy. Then the improved model is used to generate bounding boxes and segmentation mask of the cup and the oil. Finally, morphological operations are applied to the mask and oil level is computed based on pixel ratio of these two kinds of mask. Respirator images collected from substations are used to conduct experiments and the results show that the proposed method is effective in practical application.