In the field of ear shape clustering, 3D ear modeling and personal products related to ears, it is of great significance to obtain some key physiological curves of ears and the accurate positions of key points. Traditional edge extraction methods are very sensitive to changes in lighting and posture. A method for extracting key physiological curves of ears based on an improved instance segmentation network YOLACT is proposed. By labeling the ear data set, an improved YOLACT model can be trained, and the proposed strategy can be used in the prediction stage, which can accurately segment different regions of the ear and extract key physiological curves. It shows higher segmentation accuracy on the test data set, and can extract more accurate ear physiological curves. The proposed method also shows robustness under posture changes.