Module-based multiple feature integration descriptor for image retrieval

Combining various features offers distinct advantages by capitalizing on their complementary attributes to provide more refined image representations. However, effectively integrating these heterogeneous features and harnessing their strengths remains challenging. To address this, we propose a novel unsupervised approach termed Module-based Multiple Feature Integration Descriptor (MMFID) for image retrieval, drawing inspiration from the information processing mechanisms observed in the biological visual cortex. This method systematically and efficiently integrates various features within a unified framework. Within this framework, we propose an adaptive luminance mask tailored for deep structural and semantic features. It facilitates the spatial fusion of global visual cues with deep features. Additionally, we propose two parallel branches to enhance feature distinctiveness: one focuses on saliency region enhancement, while the other emphasizes semantic information integration. To achieve seamless integration, we devise a multi-module fusion strategy that harmonizes classical visual features with deep structural and semantic features. This strategy effectively exploits the complementary nature of these diverse features. Comprehensive experimental results demonstrate the competitive performance of our MMFID method across five benchmark retrieval datasets, surpassing several state-of-the-art methods based on pre-trained models.