Adaptive Two‐Tier Deep Learning for Content‐Based Image Retrieval and Classification With Dynamic Similarity Fusion

ABSTRACT Content‐Based Image Retrieval (CBIR) systems have difficulties with computing efficiency, illumination robustness and noise sensitivity. Traditional methods rely on handcrafted features or monolithic deep learning architectures, which either lack adaptability to diverse image domains or suffer from high computational complexity. To bridge this gap, a unique two‐tier deep learning system is presented in this research to overcome these drawbacks. First, a supervised neural network (SNN) reduces dimensionality and improves interpretability by converting HSV colour space into semantic 2D colour labels through pixel‐level classification. This addresses the inefficiency of processing raw RGB data while preserving illumination‐invariant colour semantics. Second, a Convolutional Neural Network (CNN) greatly increases computing efficiency by processing these labels rather than raw images. By operating on compressed 2D representations, the system achieves faster inference compared to standard 3D CNN pipelines. The framework presents Variable Weight Overall Similarity (VWOS), a versatile similarity metric that combines semantic (softmax) and structural (MaxPool3) elements with dynamically predicted weights using a neural network to automatically optimise retrieval performance based on image content. This adaptive fusion resolves the limitations of fixed‐weight similarity measures in handling heterogeneous query types. The system has achieved a performance with precision@10 scores of 0.9‐1.0 and classification accuracies of 0.85‐0.98 when tested on the PH 2 , Oxford Flowers, Corel‐1k, Caltech‐101 and Kvasir datasets. Notably, it outperforms current handcrafted, deep learning and hybrid approaches, achieving 1.0 precision@10 on four datasets and 0.96 accuracy on medical Kvasir images. Quantitative comparisons show 9%–14% higher precision than handcrafted methods, 3%–35% improvement over deep learning baselines, and 12% better than hybrid systems. This approach is especially promising for applications involving multimedia retrieval and medical imaging, where interpretability and accuracy are crucial.