Ellipse‐<scp>YOLO</scp>: A Near Real‐Time Detection of Circular Marks' Centers Network Based on High‐Speed Photogrammetry Images

ABSTRACT Circular markers are widely utilized as control points in high‐speed images measurement and industrial inspection. Accurate identification of the markers and their coordinates is crucial for calibration and structural displacement tracking. The traditional methods for circular marks' center extraction suffer from the problem of low accuracy. Although deep learning methods exhibit higher accuracy, they are usually used in object detection or contour extraction. To extract the marks' centers, traditional methods still need to be integrated on a sub‐basis. There is still a lack of an end‐to‐end circular markers' center extraction method at present. To solve these problem, this paper propose a novel deep learning based model, named Ellipse‐YOLO, for near real‐time recognizing circular markers. Based on the commonly used YOLOV8 model, three crucial improvements are made: an enhanced angle encoding prediction head, a multi‐scale oriented fusion attention module, and a novel small targets detection module, which are used to solve the discontinuous encoding of angle encoding, enhance the multi‐level feature integration, and improve the detection accuracy of small circular marks, respectively. The method has been tested in complex scenarios with varying lighting, angles, and disturbances. Experimental results demonstrate that Ellipse‐YOLO demonstrates superior performance and generalization capabilities over state‐of‐the‐art methods, with a detection accuracy of 83.5% AP and MAE of 0.019 pixels. In addition, the recognition speed of the proposed method reaches 41.2 FPS. This study not only improves the accuracy of Circular Marks detection but also pioneers the incorporation of angle information in the research of Circular Marks' Centers recognition.