Overview
A research survey details the application and evolution of supervised deep learning architectures for infrared small target detection. The paper organizes existing literature by modeling strategies, examines the critical scenarios these networks are designed to navigate, summarizes standard performance evaluation protocols, and outlines directions for future inquiry.
Research Context
Infrared small target detection previously relied on traditional unsupervised algorithms and low-rank and sparse decomposition models, which were examined in the precursor paper, "Traditional and Recent Approaches for Infrared Small Target Detection : Overview, Challenges and Perspectives (Part 1)". Following the development and adoption of supervised deep learning, neural network-based approaches have surpassed those traditional unsupervised and low-rank decomposition frameworks.
Approach
The survey provides a structured taxonomy of deep learning models in the domain by assessing three core components:
- Modeling Strategies: Categorization of architectures deployed in the literature, specifically convolutional neural networks, generative adversarial networks (GANs), Transformers, Mixture of Experts, and graph and hypergraph neural networks.
- Critical Situations: Analysis of the challenging operational conditions that individual deep learning models claim to handle.
- Performance Protocols: Review of specific performance evaluation protocols used across the research landscape to assess model efficacy.
Findings
Deep neural network methods have demonstrated performance surpassing traditional unsupervised baselines and low-rank/sparse decomposition unsupervised models. The literature reflects a broad shift toward diverse neural network paradigms—spanning convolutional backbones, adversarial setups, attention-based Transformers, modular Mixture of Experts frameworks, and recent graph/hypergraph structures—while identifying several open directions for future research in infrared small target detection.