Abstract—The necessity for effective and efficient Content- Based Image Retrieval (CBIR) systems—where retrieval is done based on the visual content of images rather than metadata—has increased due to the exponential expansion of digital image repositories. Conventional CBIR methods that depend on manually created features frequently fail to capture intricate semantic links in image data, which restricts their scalability and retrieval per- formance. Recent research has investigated deep learning-based representations and hashing methods for increased efficiency and accuracy in order to overcome these issues. Nevertheless, a lot of current methods either need a lot of labeled data or are unable to produce compact representations that are quick to retrieve in large-scale environments. In this study, we offer a deep unsupervised CBIR framework that combines semantic hashing with convolutional autoencoders. A semantic hashing mechanism is used to convert the discriminative, low-dimensional feature embeddings that the autoencoder learns into compact binary hash codes. This makes it possible to use the Hamming distance for an effective similarity search. Comprehensive tests on benchmark datasets like Caltech-101 and CIFAR-10 show that our method provides a scalable and efficient solution for large-scale image retrieval problems by achieving competitive retrieval accuracy while drastically cutting down on retrieval time.
Index Terms—Content-Based Image Retrieval, Deep Autoen- coder, Semantic Hashing, Image Embedding, Binary Hash Codes
