Multi-focus image fusion method

Multi-focus image fusion based on pre-trained sparse coding and guided filter decision maps

This technology is a multi-focus image fusion method that determines focus regions by calculating sparse coefficients using a dictionary learned from multiple images of the same scene with different focal points, then refines these into a decision map to merge them into a single, fully focused image.

Surveillance, traffic, and industrial vision sensor systems often struggle to capture a fully focused scene in a single image due to limited depth of field. This complicates image analysis and scene understanding, while also leading to redundant storage requirements from having to save multiple images with different focal points.

This technology learns a 64×512 dictionary using patches blurred by a Gaussian filter. It then generates a final decision map with smooth boundaries by removing non-smooth regions from an initial map, upsampling, and applying guided filtering. Applicable to urban security cameras and industrial machine vision, it produces extended depth-of-field images with high contrast and minimal information loss.

‍

‍

Key Features:

• Learning a dictionary by extracting random patches from input images with different focal points after blurring focus and non-focus areas with a Gaussian filter
• Calculating sparse coefficients using the learned dictionary and generating score maps and initial decision maps for each image via sliding window sampling
• Generating a final decision map by removing non-smooth regions from the initial map, upsampling to the input size, and applying guided filtering
• Setting the dictionary size to 64×512 and patch size to 8×8, then fusing the multi-focus images based on the final decision map

‍