In object detection tasks, the detection accuracy and recall rate of small size are not satisfied. This paper proposes a multi-pass object detection algorithm, applied in flaw inspection of electric metering devices. The images in the training set are categorized by clustering the labeled positions. In the first detection pass, the input image is inspected by the YOLO method, and then the parts of the image are cropped according to the type of the image and detected by a second pass. Experiments show that multi-pass detection is superior to a single pass of detection, and achieves 4% higher average recall rate than single-pass detection.