Reinforcement Learning-Based Sequential Parameter Tuning for Image Signal Processing

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Hardware image signal processing (ISP) involves complex parameter tuning, traditionally manual and subjective.
• Existing deep learning methods often overlook intrinsic relationships between ISP modules, treating the process as a black box.

Purpose of the Study:

• To develop an automated and efficient ISP parameter optimization model.
• To explore the impact of sequential ISP module structure and parameter coupling on tuning.

Main Methods:

• Introduced a single-agent reinforcement learning (RL) model (SARL-ISP) for sequential ISP parameter optimization.
• Proposed a multi-agent RL (MARL-ISP) framework incorporating a serialized parameter tuning module (SPTM) and feature selection module (FSM).

Main Results:

• SARL-ISP and MARL-ISP models demonstrate effectiveness and efficiency across various tasks like object detection and instance segmentation.
• Models achieve superior performance compared to state-of-the-art methods, even with minimal training data.

Conclusions:

• Reinforcement learning offers a robust framework for optimizing hardware ISP parameters.
• The proposed SARL-ISP and MARL-ISP models provide significant improvements in image quality and processing efficiency.