Contrastive learning-guided hashing model for artwork image retrieval

Abstract

With the growing demand for spiritual enrichment, artwork retrieval has gained increasing popularity in computer vision. While current quick response (QR) code-based and deep learning-based methods enable real-time artwork recognition, their dependence on manual annotations limits adaptability and scalability. In this paper, we propose an unsupervised hashing-based artwork retrieval framework that requires no manual labeling and adapts to new classes. The system comprises two key components: a detection module and a retrieval module. The detection module processes user-captured artwork images by predicting polygonal bounding boxes and rectifying them through perspective transformation algorithms. The retrieval module is the focus of our research. We present a novel contrastive learning framework for artwork retrieval that integrates a convolutional neural network (CNN) based feature extractor and a hashing encoder–decoder structure. This architecture processes input images into both floating-point and hashing representations while maintaining a binary memory bank for efficient similarity matching. Two specialized loss functions facilitate adaptive hashing encoding, aligning floating-point and hashing features through an unsupervised learning process. We evaluate our framework on a self-built dataset containing over 24k images spanning traditional Chinese paintings, oil paintings, and Chinese chops. The samples are unlabeled and there is only one image of each artwork. Comparative experiments with state-of-the-art methods demonstrate our system’s superior effectiveness and strong potential for practical artwork retrieval applications.