Authors - Gokaramaiah Thota, Sathya Babu Korra, Nagaraju K, Suman Prakash, Perumalla, V Ramanjaneyulu Yannam Abstract - Content-based image retrieval (CBIR) has become an important research area in computer vision for retrieving visually similar images from large-scale image collections using intrinsic image content rather than manually assigned annotations. Although deep feature learning has substantially improved retrieval accuracy, existing retrieval systems continue to encounter challenges, including limited interpretability, high computational complexity during similarity search, and insufficient integration of semantic attention information into retrieval and indexing mechanisms. Most retrieval acceleration techniques operate primarily in feature space and often neglect region-level semantic cues that contribute to retrieval relevance. This work presents a hybrid framework that integrates complete optimization of convolutional features, explainable artificial intelligence (XAI)-guided feature representation, and histogramdriven indexing to improve retrieval effectiveness while reducing computational cost. The complete convolutional neural network architecture is optimized to learn domain-adaptive visual representations, while class activation information is employed to identify semantically important image regions and generate activation-aware feature embeddings that preserve discriminative visual characteristics. The extracted features are normalized and these are subsequently organized into histogram partitions to reduce the retrieval search space. This indexing strategy limits similarity computation to semantically related candidate regions while preserving retrieval quality. Experiments conducted on Corel-10K and Oxford-5K evaluate retrieval performance using Precision@K, average precision, mean average precision, retrieval time, and average feature comparisons. Experimental results demonstrate that the proposed framework reduces the retrieval search space by approximately 35 to 55% while maintaining competitive retrieval accuracy. Comparative analysis shows competitive retrieval performance with lower computational complexity and improved interpretability.
