Abstract:While deep learning (DL) methods based on empirical risk minimization (ERM) have achieved superior detection performance on public infrared small target detection (IRSTD) datasets, the learned models suffer severe performance degradation when dealing with out-of-distribution (OOD) data. The primary challenge in practical infrared systems lies in enhancing the generalization ability to ensure reliable detection performance in diverse scenarios. In this article, we propose a single-source domain generalization (SDG) and test-time adaptation (TTA) approach for IRSTD, aiming to train a domain-agnostic DL model in the presence of only one source domain and enabling it to perform well on any unseen target domain. During the training process, perturbed images and networks (PINs) are utilized to enhance the diversity of features extracted from a single source domain with limited samples. A Y-shaped multibranch network employs a combination of the self-supervised recovery task and the supervised detection task. Their losses jointly guide the network to converge explicitly to the flat minima in the loss landscape. During the test process, we apply domain-guided adaptation (DGA) to the learned model. The feature statistics of the dynamically inputted image batches are considered as cues to the target domain. Batch normalization (BN) layers are adaptively fine-tuned online based on the differences in feature statistics between the source model and target images. We conducted extensive ablation studies to demonstrate the rationality and effectiveness of each component of the framework. Compared to other SDG, TTA, and model-driven IRSTD methods, PIN-DGA can more effectively improve the detection performance of DL methods on unseen target domains. As a model-agnostic framework, we verified its compatibility with current state-of-the-art networks. On three OOD datasets, our method can improve the average probability of detection (PD) by 16.08 and reduce false alarm targets (FATs) by 33%. The codes are available at https://github.com/jzchenriver/PIN-DGA.

Enhanced Generalization Ability of Infrared Small Target Detection via Perturbed Training and Adaptive Test

Domain Adaptation for Infrared-Radar Cross-Scene Multimodal Detection

Prior-Guided Data Augmentation for Infrared Small Target Detection

A Semantic Domain Adaption Framework for Cross-Domain Infrared Small Target Detection

Region-guided Network with Visual Cues Correction for Infrared Small Target Detection

Guided Attention and Joint Loss for Infrared Dim Small Target Detection

Mitigate Target-level Insensitivity of Infrared Small Target Detection via Posterior Distribution Modeling

Enhanced Cross-Domain Dim and Small Infrared Target Detection via Content-Decoupled Feature Alignment

Dual-Domain Prior-Driven Deep Network for Infrared Small-Target Detection

Unleashing the Power of Generic Segmentation Model: A Simple Baseline for Infrared Small Target Detection

A Feature Enhancement and Augmentation-Based Infrared Small Target Detection Network

Edge-Guided Perceptual Network for Infrared Small Target Detection

Enhancing Infrared Small Target Detection Robustness with Bi-Level Adversarial Framework

Adaptive Domain Generalization via Online Disagreement Minimization

Dual-Stream Edge-Target Learning Network for Infrared Small Target Detection

Towards Domain Generalization in Object Detection

On the Connection Between Invariant Learning and Adversarial Training for Out-of-Distribution Generalization

An infrared small target detection model via Gather-Excite attention and normalized Wasserstein distance

MDIGCNet: Multi-Directional Information-Guided Contextual Network for Infrared Small Target Detection

FDDBA-NET: Frequency Domain Decoupling Bidirectional Interactive Attention Network for Infrared Small Target Detection

Receptive-Field and Direction Induced Attention Network for Infrared Dim Small Target Detection with a Large-Scale Dataset IRDST.