Chinese painting seal recognition based on EfficientNet and scale-invariant feature transform (SIFT)

To improve the efficiency and accuracy of Chinese painting seal recognition, this paper proposes a two-stage seal recognition algorithm based on EfficientNet and scale-invariant feature transform (SIFT). In the initial stage of seal extraction, preprocessing techniques are utilized to optimize image quality. HSV (hue, saturation, value) color space features are employed to identify potential seal regions, followed by the application of the EfficientNet model to extract image features from these candidate regions for classification and the retrieval of seal images. In the subsequent seal matching stage, the SIFT algorithm is employed to extract image features from the seal images, and nearest neighbor matching is conducted to obtain the final seal information. A dataset comprising 4000 images for seal extraction and a standard seal database consisting of 14790 records of seal information are created to evaluate the algorithm’s effectiveness. The new method achieves an accuracy of 95.25% in seal image extraction and 98.20% in seal matching using the self-built dataset. Furthermore, the method affords robust handling of image rotation and scale changes.