Abstract
Digital images have increased rapidly, and effective image retrieval systems have been developed to accommodate this growth. Handcrafted methods are not always effective at representing complex image semantics, and this has resulted in a trend towards deep learning based methods. The state-of-the-art techniques are currently used, which automatically extract high-dimensional features using convolutional neural networks (CNNs) and self-supervised learning (SSL). But, there still remain scope for improving retrieval accuracy and efficiency, especially on processing a variety of datasets. In this work, an image retrieval model is proposed which combines pre-trained EfficientNetB0 and ResNet18 backbones based on SimCLR contrastive learning for better feature learning, while a learnable fusion layer is applied to reduce the dimension of the learned features. In this layer, it learns to select and compress certain dimensions that helps the most with the learning task. Learn adaptive weights, which automatically focus on important dimensions from both backbones. The model gets features from all the images and keeps them as feature vectors in a MongoDB database, enabling scalable and efficient search. After inputting a query image (QI), the model extracts its features from the trained hybrid model and uses Facebook AI Similarity Search (FAISS) to quickly retrieve the most similar features from the stored embeddings. The number of most similar feature images is retrieved from the database and shown in the web interface. In the case of self-supervised approaches, this research uses various benchmark datasets to train the hybrid model to learn useful representations without needing labelled data. The model shows an average accuracy rate of 97% through experimentation, indicating significant improvements. A learnable fusion layer, hybrid feature extraction, self-supervised learning, and scalable database management have proven the system's ability to deliver accurate and efficient image retrieval. This work illustrates the combination of multimodal CNN backbones and contrastive learning for state-of-the-art content-based image retrieval (CBIR) solutions, applicable to a wide variety of and large-scale image datasets.
Replies