Abstract
Content-based image retrieval (CBIR) systems have formidable obstacles in connecting human comprehension with machine-driven feature extraction due to the exponential expansion of visual data across many areas. Robust performance across varied datasets is challenging for traditional CBIR methods due to their reliance on hand-crafted features and inflexible structures. This study presents a deep adaptive attention network (DAAN) for CBIR that combines multi-scale feature extraction and hybrid neural architectures to solve these problems and improve the speed and accuracy of visual retrieval. The DAAN architecture integrates transformer-based models for capturing picture contextual connections with deep neural network (DNN) to extract spatial features. A new adaptive multi-level attention module (AMLA) that guarantees accurate feature weighting improves the system's ability to detect minute visual material changes. Findings show that DAAN-CBIR outperforms existing approaches with high mean average precision (map), retrieval speed, and reduced training time. These developments prove its efficacy in various fields, including e-commerce, digital information preservation, medical imaging diagnostics, and personalized media recommendations.