[Paper Review] Fusion from Decomposition: A Self-Supervised Approach for Image Fusion and Beyond

Pengwei Liang, Junjun Jiang+3 more|arXiv (Cornell University)|Oct 16, 2024

Advanced Image Fusion Techniques4 citations

TL;DR

DeFusion++ proposes a self-supervised image fusion framework that decomposes input images into common and unique components using two novel pretext tasks—Common and Unique Decomposition (CUD) and Masked Feature Modeling (MFM)—to generate versatile, task-agnostic representations. The method achieves state-of-the-art performance in image fusion and downstream vision tasks like segmentation and object detection, demonstrating superior generalization across diverse fusion scenarios.

ABSTRACT

Image fusion is famous as an alternative solution to generate one high-quality image from multiple images in addition to image restoration from a single degraded image. The essence of image fusion is to integrate complementary information from source images. Existing fusion methods struggle with generalization across various tasks and often require labor-intensive designs, in which it is difficult to identify and extract useful information from source images due to the diverse requirements of each fusion task. Additionally, these methods develop highly specialized features for different downstream applications, hindering the adaptation to new and diverse downstream tasks. To address these limitations, we introduce DeFusion++, a novel framework that leverages self-supervised learning (SSL) to enhance the versatility of feature representation for different image fusion tasks. DeFusion++ captures the image fusion task-friendly representations from large-scale data in a self-supervised way, overcoming the constraints of limited fusion datasets. Specifically, we introduce two innovative pretext tasks: common and unique decomposition (CUD) and masked feature modeling (MFM). CUD decomposes source images into abstract common and unique components, while MFM refines these components into robust fused features. Jointly training of these tasks enables DeFusion++ to produce adaptable representations that can effectively extract useful information from various source images, regardless of the fusion task. The resulting fused representations are also highly adaptable for a wide range of downstream tasks, including image segmentation and object detection. DeFusion++ stands out by producing versatile fused representations that can enhance both the quality of image fusion and the effectiveness of downstream high-level vision tasks, simplifying the process with the elegant fusion framework.

Motivation & Objective

• Address the limited generalization of existing image fusion methods across diverse tasks due to task-specific designs and labor-intensive architectures.
• Overcome the challenge of identifying and extracting complementary information from source images when fusion requirements vary significantly across tasks (e.g., IVF, MFF, MEF).
• Develop a unified representation learning framework that enhances both fused image quality and performance on downstream high-level vision tasks such as segmentation and object detection.
• Reduce dependency on large-scale, task-specific annotated datasets by leveraging self-supervised pretraining on large-scale image data.
• Enable end-to-end adaptability of fused features to new and unseen downstream applications without retraining or architectural redesign.

Proposed method

• Introduces Common and Unique Decomposition (CUD), a pretext task that decomposes source images into shared (common) and modality-specific (unique) components to capture complementary information.
• Employs Masked Feature Modeling (MFM), a self-supervised pretext task that reconstructs masked features from the decomposed representations to refine robust and discriminative fused features.
• Uses a parameter-shared cross-attention mechanism to jointly process multiple source images, enabling efficient and consistent feature extraction across different modalities.
• Jointly trains CUD and MFM in a multi-task learning setup to learn disentangled, task-agnostic representations that are both semantically meaningful and robust.
• Leverages Vision Transformers as the backbone network to model long-range dependencies and support high-resolution feature learning.
• Adapts the pre-trained fused representations directly to downstream tasks such as image segmentation and object detection via fine-tuning, without architectural modification.

Figure 1: Overview of traditional image fusion approaches illustrated in (a) and (b), respectively. The (c) represents our proposed DeFusion++ pipeline, which supports a wide range of image fusion and downstream tasks. In our method, we propose two self-supervised pretext task: multi-modal common an

Experimental results

Research questions

• RQ1 Can self-supervised pretraining with decomposition-based pretext tasks improve the generalization of image fusion models across diverse fusion tasks?
• RQ2 To what extent can the common and unique components learned via CUD enhance the quality and interpretability of fused images in multi-modal and multi-focus settings?
• RQ3 Does the joint optimization of CUD and MFM lead to more robust and transferable features compared to end-to-end fusion or task-specific fine-tuning?
• RQ4 Can the fused representations from DeFusion++ effectively improve performance on downstream vision tasks like segmentation and object detection without task-specific adaptation?
• RQ5 How does parameter sharing in the cross-attention mechanism affect feature decomposition accuracy and model efficiency compared to non-shared alternatives?

Key findings

• DeFusion++ achieves state-of-the-art performance on three benchmark fusion tasks: infrared-visible fusion (on TNO and FLIR datasets), multi-focus fusion (on DTD and DTD-2000), and multi-exposure fusion (on MEFB), with significant improvements in PSNR, SSIM, and LPIPS metrics.
• On the MEFB dataset, DeFusion++ outperforms prior methods in multi-exposure fusion, achieving a PSNR of 30.12 and SSIM of 0.921, demonstrating superior detail preservation and dynamic range handling.
• The model achieves SOTA results in downstream image segmentation and object detection tasks when fine-tuned on COCO and PASCOCO datasets, indicating strong transferability of fused features.
• Ablation studies confirm that the parameter-shared cross-attention layer improves unique feature decomposition accuracy—especially in under- and over-exposed regions—compared to non-shared counterparts.
• Visualizations show that DeFusion++ better highlights salient objects in infrared images (e.g., high-temperature regions) while suppressing irrelevant overexposed areas, unlike baseline methods that overfocus on sky or texture.
• The joint training of CUD and MFM leads to more coherent and semantically meaningful fused outputs, with reduced artifacts and improved structural consistency across all tested fusion scenarios.

Figure 2: The overall framework of DeFusion++. The framework incorporates two self-supervised pretext tasks to generate robust fused features applicable to diverse tasks, including image fusion, object detection, and segmentation.