Claim Missing Document
Check
Articles

Found 1 Documents
Search
Journal : Journal of Intelligent Computing and Health Informatics (JICHI)

Evaluation of a Semantic Representation-Based Retrieval Model on a Text Dataset Generated from Image Transformation Firmansyah, Muhammad; Marutho, Dhendra; Ilham, Ahmad; Saputra, Irwansyah
Journal of Intelligent Computing & Health Informatics Vol 6, No 2 (2025): September
Publisher : Universitas Muhammadiyah Semarang Press

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.26714/jichi.v6i2.19240

Abstract

The increasing demand for efficient multimodal information retrieval has driven significant research into bridging visual and textual data. While sophisticated models like CLIP offer state-of-the-art semantic alignment, their substantial computational requirements present challenges for deployment in resource-constrained environments. This study introduces a lightweight retrieval framework that leverages the BLIP image captioning model to transform image data into rich textual descriptions, effectively reframing cross-modal retrieval as a text-to-text task. We systematically evaluated three retrieval models BM25, SBERT, and T5 on caption-transformed MSCOCO and Flickr30K datasets, utilizing both classical metrics (Recall@5, mAP) and semantic-aware metrics (SAR@5, Semantic mAP). Experimental results demonstrate that T5 achieves superior semantic performance (SAR@5 = 0.561, Semantic mAP = 0.524), surpassing SBERT (SAR@5 = 0.524) and outperforming the lexical BM25 baseline (SAR@5 = 0.312). Notably, the proposed BLIP+T5 pipeline attains 88% of CLIP’s semantic accuracy while reducing inference latency by approximately 60% and decreasing GPU memory consumption by over 60%. These findings underscore the potential of caption-based retrieval frameworks as scalable, cost-effective alternatives to computationally intensive multimodal systems, especially in latency-sensitive and resource-limited scenarios. Future work will explore fine-tuning strategies, domain-adapted semantic metrics, and robustness under real-world conditions to further advance retrieval effectiveness.
Co-Authors Abiyah, Indra Adib Adhitya Bagus Singandaru Adibrata, Sudirman Ainul Yaqin Ali Rahman Amrillah Apriani, Isna Apriyandi, Putra Arif Maulana, Arif Aslimah, Aslimah Aulia, Nur Afti Az Zahra, Alifia Bhaskara, Renaldi Surya Chairul Abdi Dhendra Marutho Dyo Arga Nata, Muhamad EMI SALMAH Fikri, Sultoni Firdausy, Muhammad Abrar Firmansyah, Esa Frans Simangunsong Gunawan, Bambang Ady hartanto hartanto Hendarti, Hendarti Herdinov, Adam Hidayat, Lutfi Alvian Hikmawati, Inayah Himawan Sutanto, Himawan Ilham, Ahmad Imroatul Khasanah Irhash, Irhash Irwan Irwan Iwan Harsono Jati, Dian Rahayu KARISMAWAN, PUTU Khairunisa, Assavira M. Irwan M. Irwan, M. Irwan Maria, Linda Maulana Al Aby, Theo Maulidia, Vera Mizwar, Andi MUAIDY YASIN, MUAIDY Muftiadi, M. Rizza Muhammad Agreindra Helmiawan Muhammad Aria Wahyudi Muhammad Sri Wahyudi Suliswanto Musyoliha, Tuti Muzakar, Muzakar Mu’ammal, Immanuel Nilawati, Indah Nirtha Nofitasari, Solehati Nopi Stiyati Prihatini Nugraha, Nanda Nur, Alqadri Nuraulia, Nasyah Octavia, Azella Nosih Paramita Sari, Paramita Perangin-angin, Robet Prabowo, Ien Yus Rizal Purnama, Isnan Purwanti, Irna Putra, Surya Dwi Rijal, Muhammad Nor Riza Miftahul Khair Rizqi Puteri Mahyudin SABINA, VINA Sadiq, Ahmad Safitri, Nur Aprilia Sahri Sahri Sahri Saputra, Irwansyah Sari, Nova Puspita Septiyani, Santi Sikin, Muhamad St. Maryam, St. Suphia, Suphia T, Thoyyibah Tanjung , Yanto Ul Walidaien, Syifa Ulfa, Yakut Nahdiana Wahidin Wahidin Wahyuni , Hera Widari, Baiq Widyanor, Julian Zara Tania Rahmadi