Research on Lightweight Image Matching and Cattle Individual Identification Technology Based on Subpixel Keypoints

To enhance the recognition model's adaptability to non-standardized data, this study proposes a lightweight image matching-based cattle individual identification technique that utilizes subpixel keypoints. The technique leverages SuperPoint and LightGlue to construct an image matching algorithm, with improvements made to enhance recognition accuracy. During the feature point extraction process, keypoint refinement is introduced, using the learned displacement vectors of features to enhance SuperPoint's subpixel accuracy. Additionally, the OTSU algorithm is employed to compute the feature extraction threshold adaptively, improving the feature point extraction process. A two-layer validation screening method is employed to optimize the matching pairs of LightGlue, further improving matching efficiency. To validate the effectiveness of the algorithm, comparative experiments were conducted on a self-constructed cattle facial dataset, comparing it with various image matching methods. The results indicate that, on the narrow-baseline dataset, the macro-average precision, recall, and F1 scores are 97.87%, 97.50%, and 97.68%, respectively. On the wide-baseline dataset, these metrics are 85.09%, 74.75%, and 79.59%, respectively. All results significantly surpass those of traditional image matching algorithms. In conclusion, the image matching algorithm proposed in this study effectively improves the cattle individual recognition model's adaptability to non-standardized data, providing valuable technical references for the practical application of cattle individual recognition methods.