Structural Similarity Loss for Learning to Fuse Multi-Focus Images

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Existing multi-focus image fusion methods often rely on simulated data and complex post-processing.
• Supervised learning approaches require ground truth images, which are difficult to obtain for fusion tasks.

Purpose of the Study:

• To develop an end-to-end deep learning model for direct multi-focus image fusion.
• To eliminate the need for ground truth data and complex post-processing steps in image fusion.

Main Methods:

• A convolutional neural network (CNN) architecture was designed for direct fusion of multi-focus images.
• The model utilizes image structural similarity (SSIM) and local window standard deviation for loss calculation, avoiding ground truth data.
• The fully convolutional network handles variable image sizes, enabling training on real benchmark datasets.

Main Results:

• The proposed method achieves direct fusion without simulated data or ground truth.
• The network successfully fuses multi-focus images using SSIM and standard deviation in the loss function.
• Performance evaluations demonstrate that the method is comparable or superior to state-of-the-art techniques.

Conclusions:

• The end-to-end learning approach offers an effective solution for multi-focus image fusion.
• The method's ability to use real datasets and variable image sizes enhances its practical applicability.
• This work advances direct image fusion techniques by leveraging unsupervised learning principles.