DroneIQA-VLE: Multi-Task Drone Image Quality Assessment via Vision-Language Ensemble

2026-07-01

AI Summary

This work addresses the challenge of simultaneously evaluating image quality across global, target, and background regions in low-altitude drone imagery by proposing a dual-path multi-task image quality assessment method. One path leverages a SigLIP2 vision encoder with a multi-task regression head, while the other employs a LoRA-finetuned Qwen3.5-9B multimodal large language model for quality regression. The final global quality prediction is obtained by arithmetically averaging the outputs of both pathways. To the best of our knowledge, this is the first approach to integrate a vision–language foundation model with a dedicated vision encoder for drone image quality assessment, enabling joint prediction of multi-region quality. The method achieved second place in the ICME 2026 Drone-IQA Grand Challenge, demonstrating its effectiveness and state-of-the-art performance.

Abstract

We present DroneIQA-VLE, our solution to the ICME 2026 Drone-IQA Grand Challenge on Target-aware Image Quality Assessment for Low-altitude UAV Images. The framework jointly predicts global, target, and background quality scores by ensembling two complementary pipelines: (1) SigLIP2 vision encoders with multi-task regression heads, and (2) a LoRA-adapted Qwen3.5-9B multimodal large language model for quality score regression. The final global quality prediction is obtained by arithmetically averaging the outputs of both pipelines. Our method achieves 2nd place in the challenge, demonstrating its effectiveness. The code is available at https://github.com/sunwei925/DroneIQA-VLE.

Problem

• Drone Image Quality Assessment
• Target-aware IQA
• Low-altitude UAV Images
• Multi-task IQA

Innovation

• vision-language ensemble
• multi-task image quality assessment
• LoRA-adapted LLM
• SigLIP2 encoder
• UAV image quality