Content-based image retrieval based on luminance and texture decomposition

Abstract The main challenge in content-based image retrieval (CBIR) systems is accurately describing image features in a manner that aligns with human perception. Natural images consist of varying luminance levels and complex, intertwined texture patterns. Decomposing an image into its constituent components can simplify feature extraction and enhance system performance. This paper proposes a method for CBIR based on luminance and texture decomposition. In the luminance component, texture features are suppressed, allowing luminance features to be extracted using AlexNet. Conversely, in the texture component, texture features are amplified and extracted using SqueezeNet. AlexNet captures the global context of images by utilizing spatial information, thereby enhancing feature contrast for better discrimination between object classes. SqueezeNet is selected for its ability to produce compact, highly discriminative feature vectors that effectively describe texture features. To mitigate redundancy caused by feature overlap, the most relevant features are selected using the Boruta–Shap algorithm. The feature space is visualized using the t-distributed stochastic neighbor embedding (t-SNE) technique, and the interpretability of the proposed approach is evaluated through Shapley value analysis. Experimental results demonstrate the effectiveness of the proposed approach in improving CBIR performance.