.png){kind=icon}
.png){kind=icon}
Return
HA-ISP: Towards Human-Aligned automatic ISP hyperparameter tuning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2026.132282.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Image Signal Processing (ISP) transforms RAW sensor data into high-quality images that meet diverse downstream requirements through the adjustment of abstract algorithm parameters in the ISP pipeline. Recent research has explored neural network proxy models and reinforcement learning (RL) for automatic optimization in high-dimensional, nonlinear ISP parameter spaces, but existing methods often suffer from complex architectures with limited generalization, biased reward signals, and incomplete iterative strategies during both training and inference, which collectively restrict their deployment in real ISP systems. To address these challenges, we propose an automated ISP parameter optimization framework based on multimodal large language models (MLLMs). For consumer camera ISPs, we design a three-stage training paradigm comprising supervised fine-tuning, rejection sampling–based data filtering, and reinforcement learning, coupled with a multi-dimensional image quality assessment model as a unified feedback signal to improve data efficiency, simplify the training pipeline, and achieve tuning results that better align with human subjective perception. Furthermore, we introduce a retrieval-augmented generation (RAG)–based memory-augmented iterative reasoning mechanism that leverages historical information during inference to significantly enhance tuning efficiency and stability. Overall, our method establishes an end-to-end ISP tuning paradigm that is automated and human-aligned for consumer cameras, and with only lightweight natural language adaptation, it can be efficiently transferred to other downstream vision tasks while maintaining strong performance.
Keywords:
ISP optimization
multimodal large language models
reinforcement learning
image quality assessment
retrieval-augmented generation
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W
Organization
No organization information available