A Scheme to Improve the Detection Accuracy of Small Parts of Electric Meters.

It is a challenging task to detect small parts of the ammeter, which is essential to defect inspection on meters. In this paper, we propose a multi-pass detection method to improve the accuracy of detecting small objects. The position distribution patterns of small parts are found by clustering the data of the relative position to the meter in the labeled dataset. In the first pass, the whole meter is detected. In the second pass, the potential areas including small parts are found according to the distribution patterns, and then they are cropped and sent to the detection process to find out small objects. The experimental results show that the method can effectively improve the accuracy of small part detection, compared with object detection without utilizing distribution pattern of small objects.