PERFORMANCE COMPARISON OF VISION-LANGUAGE MODELS IN IMAGE CLASSIFICATION

Authors : Doğukan Özeren, Erkan Yüksel, Asım Sinan Yüksel

Pages : 247-262

Doi:10.46519/ij3dptdi.1693015

View : 82 | Download : 191

Publication Date : 2025-08-30

Article Type : Research Paper

Abstract :Vision-Language Models (VLMs) have introduced a new paradigm shift in image classification by integrating visual and textual modalities. While these models have demonstrated strong performance on multimodal tasks, their effectiveness in purely visual classification remains underexplored. This study presents a comprehensive, metric-driven comparative analysis of eight state-of-the-art VLMs—GPT-4o-latest, GPT-4o-mini, Gemini-flash-1.5-8b, LLaMA-3.2-90B-vision-instruct, Grok-2-vision-1212, Qwen2.5-vl-7b-instruct, Claude-3.5-sonnet, and Pixtral-large-2411—across four datasets: CIFAR-10, ImageNet, COCO, and the domain-specific New Plant Diseases dataset. Model performance was evaluated using accuracy, precision, recall, F1-score, and robustness under zero-shot and few-shot settings. Quantitative results indicate that GPT-4o-latest consistently achieves the highest performance on typical benchmarks (accuracy: 0.91, F1-score: 0.91 on CIFAR-10), substantially surpassing lightweight models such as Pixtral-large-2411 (accuracy: 0.13, F1-score: 0.13). Near-perfect results on ImageNet and COCO likely reflect pre-training overlap, whereas notable performance degradation on the New Plant Diseases dataset underscores domain adaptation challenges. Our findings emphasize the need for robust, parameter-efficient, and domain-adaptive fine-tuning strategies to advance VLMs in real-world image classification.
Keywords : Vision-Language Models, Image Classification, Multimodal Learning, Zero-Shot Classification, Few-Shot Learning, Model Generalization.

ORIGINAL ARTICLE URL