Content-Based Image Retrieval Using Self-Supervised Contrastive Learning with Hybrid Backbone

Abstract

The exponential growth of digital image collections has driven the requirement for more efficient content-based image retrieval (CBIR) systems. Traditional methods relying on handcrafted features often fail to capture complex image semantics, hence shifting the paradigm towards utilizing convolutional neural network (CNN) based solutions along with self-supervised learning (SSL) which automatically extracts high-dimensional features. However, when dealing with vast and diverse datasets, retrieval performance is still not at its best. In this study, a hybrid CBIR model that integrates pre-trained EfficientNet-B0 and ResNet18 backbones within a SimCLR-based contrastive learning framework to improve feature extraction. A learnable fusion layer is introduced to adaptively compress and emphasize the most informative dimensions across both networks. Extracted embeddings from all images are stored in a MongoDB database, allowing scalable and efficient search. Upon receiving a query image, the system uses the trained hybrid model to extract its features and performs a fast similarity search using Facebook AI Similarity Search (FAISS). Retrieved results are displayed via a web interface. The proposed model is performing well in terms of efficiency and effectiveness, scoring a mAP of 96% for the custom multi-class dataset and 98% for the Corel-1K benchmark dataset following training on them without labels. The combination of hybrid backbone fusion, self-supervised contrastive learning and scalable database management clearly reflects the ability of the system to provide image search efficiently.