IIT PATNA CV 1 at Sci-ImageMiner 2026 Task 1: LoRA-Enhanced Generative Vision–Language Models for Scientific Figure Panel Classification

Abstract

This paper presents a solution to the Sci-ImageMiner Competition: Task 1 (classification) at ICDAR 2026. The objective is to accurately classify individual scientific figure panels from Atomic Layer Deposition and Etching (ALD/E) literature into a domain-specific taxonomy. We formulate this challenge as a conditional text generation problem rather than a traditional discriminative classification task. By employing Low Rank Adaptation (LoRA), we efficiently fine-tune the Qwen2.5-VL-7B-Instruct vision language model to directly generate taxonomy-aware class labels in natural language, conditioned on cropped panel images and a taxonomy-aware structured prompt. This generative approach retains the model’s pre-trained scientific visual reasoning capabilities without requiring architectural modifications. Our proposed system demonstrates highly competitive performance, achieving a micro F1 score of 0.77 and securing 3rd place on the official evaluation leaderboard.