EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

A Hybrid Deep Learning Vision Language Framework for a Robust Image Deduplication.

Bollipelly Pruthvi Raj Goud

Journal of Intelligent Decision Making and Information Science · 01.10.2026 · 3 (15s) · 256-285

DOI: 10.17762/jidmis.v3.4513 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Intelligent Decision Making and Information Science, online erschienen am 01.10.2026. Verfasst von Bollipelly Pruthvi Raj Goud. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

With the growing volume of data repositories, ensuring data quality becomes a major concern. One of the problems faced by researchers is that of duplicate data. It arises due to the integration of various data sources. The process of image data deduplication is essential for avoiding redundancy in storage, managing data efficiently and retrieving images in large scale image repositories. In recent times, deep neural networks (DNNs) and vision-language models (VLMs) have shown promising performance in reducing the impact of data duplication in deployment environments. However, existing deep learning approaches often fail to simultaneously capture fine-grained visual characteristics, hierarchical spatial dependencies, and high-level semantic relationships required for robust near-duplicate detection. To overcome these challenges, this research proposes a hybrid deep Learning–Vision-Language Framework with Adaptive Feature Fusion for Robust Image Deduplication. The proposed framework employs a Siamese architecture with shared parameters, integrating an attention-enhanced EfficientNetV2-S branch for discriminative local visual representation, a hierarchical Swin Transformer V2 branch with patch partitioning, patch merging, and three transformer blocks for spatial-contextual representation, and a CLIP-based vision-language encoder for high-level semantic feature extraction. The heterogeneous representations are projected into a common latent space and integrated using an Adaptive Feature Fusion (AFF) mechanism that dynamically assigns weights to local, global, and semantic features. The resulting fused embedding are optimized using contrastive learning, while cosine similarity is used to measure correspondence between image pairs and distinguish duplicate from non-duplicate instances. Comprehensive experimentation has been conducted using two datasets, INRIA Holidays and UKBench, and performance metrics such as accuracy, precision, and recall have been analyzed. Furthermore, Grad-CAM is applied to provide interpretability, indicating the duplicate regions that aid in better detection. Results demonstrate that the proposed deduplication process has achieved an accuracy of 0.989, precision of 0.987, recall of 0.986, specificity of 0.99 and F1-score of 0.98, showing its promising performance in image data deduplication.

Zitieren

APA

Goud, B. P. R. (2026). A Hybrid Deep Learning Vision Language Framework for a Robust Image Deduplication. Journal of Intelligent Decision Making and Information Science, 3(15s), 256-285. https://doi.org/10.17762/jidmis.v3.4513

Vancouver

Goud BPR. A Hybrid Deep Learning Vision Language Framework for a Robust Image Deduplication. Journal of Intelligent Decision Making and Information Science. 2026;3(15s):256-285. doi:10.17762/jidmis.v3.4513

BibTeX
@article{Goud2026,
  author = {Goud, Bollipelly Pruthvi Raj},
  title = {A Hybrid Deep Learning Vision Language Framework for a Robust Image Deduplication},
  journal = {Journal of Intelligent Decision Making and Information Science},
  year = {2026},
  volume = {3},
  number = {15s},
  pages = {256--285},
  doi = {10.17762/jidmis.v3.4513}
}
Kostenlos starten →