Learn to Fuse: A Reinforcement Learning Framework for Multi-Focus Image Fusion

Multi-Focus Image Fusion is a computational technique used to merge multiple images of the same scene, each captured with different focus settings into a single image.This process aims to overcome the limitations of imaging systems where the depth of field is insufficient to simultaneously bring all objects in a scene into focus.This study proposes an RL-based approach for MFIF, using a patch-wise selection mechanism.It is designed to combine complementary focus information from two source images into a single and well-focused output. The system treats the fusion process as a sequential decision-making task.Here an agent determines the optimal patch selection from each source image.The agent learns a policy network, which uses Q-learning with function approximation to determine actions. The environment processes images in discrete patches and the agent selects patches based on focus measures to construct the fused image. A SSIM based reward guides the learning process. This encourages the decisions that improve the perceptual quality of the fused image. By using the RL method, this approach adapts dynamically to variations in image content thereby enabling robust fusion performance across diverse scenarios. The different datasets utilized are Lytro and MFFW. The model is implemented and evaluated on grayscale multi-focus image pairs demonstrating its effectiveness in achieving high-quality fusion results.