IndexCLR: Class-Aware Contrastive Learning for Similarity Indexing

Similarity metrics and indexing are critical components of image similarity searches since they directly impact the accuracy and efficiency of retrieving similar imagery. Similarity measurements, such as Euclidean distance or cosine similarity, evaluate how closely images are compared based on simple differences in their features. More complex methods, such as contrastive loss, enhance the representation of images by emphasizing relationships in a contextual reference to other related samples. This paper introduces a framework for image similarity and indexing that utilizes class-based representations in contrastive learning to enhance retrieval. This framework uses class-based representations for context-related awareness instead of sample-level contrastive learning to refine feature representations using a contrastive learning approach. This approach reduces computational complexity and mitigates sample selection challenges while refining the similarity comparison operation of similar images using class contextual information. A bottleneck mechanism is introduced to compress high-dimensional feature spaces into compact, lower-dimensional embeddings, preserving critical semantic information while minimizing redundancy. Discrimination is further enhanced through class-based contrastive training.