Multi-Algorithm Deep Unfolding With Application to Foreground Detection

Mustafa Siddiqui

doi:10.1109/LSP.2025.3618508

Multi-Algorithm Deep Unfolding With Application to Foreground Detection

2025-01-01

0

PRE

AI

DOI：10.1109/LSP.2025.3618508

Original

Original request for help

Save

Abstract

En 中文

Deep learning models have achieved impressive results across various image processing and computer vision tasks. However, they often require large datasets, lack transparency, and struggle to incorporate expert knowledge. Deep unfolding offers a solution for these issues by embedding traditional algorithmic steps into neural network architectures, leading to models that are both more interpretable and efficient with data. Traditionally, deep unfolding involves unrolling a single optimization algorithm into a network structure. While this approach enhances interpretability, it can limit the model’s ability to capture complex data patterns due to its constrained structure. Recent methods have tried to overcome this by adding more parameters or integrating unfolded models into larger, less interpretable systems. However, these strategies often compromise the clarity and structured design that make deep unfolding appealing. In this letter, we introduce Multi-Algorithm Deep Unfolding (MADU), a novel framework that unrolls multiple algorithmic structures within a single model. This design allows the model to learn how to balance different algorithmic approaches, enhancing flexibility while preserving interpretability and structure. We demonstrate the effectiveness of MADU on the task of foreground detection, showing that our method improves performance through a structured expansion of the model.

Keywords:

Deep learning

deep unfolding

foreground detection

structured neural networks

interpretable models

Multi-Algorithm Deep Unfolding With Application to Foreground Detection

Abstract

Journal

Organization

Cited Papers