Deep Learning for Single-Frame Infrared Small and Dim Target Detection: A Paradigm-Oriented Review with Cross-Architecture Benchmarking

Abstract

Single-frame infrared small and dim target detection underpins military early warning and space surveillance. However, existing review literature lacks a classification system based on architectural paradigms and standardized cross-architecture benchmarking. This paper systematically reviews deep learning methods from two complementary perspectives: classification based on architectural paradigms and quantitative empirical evaluation. Within our proposed analytical framework, we categorize existing methods into seven architectural paradigms—detector-based, U-Net-based, GAN-based, Transformer-based, diffusion-based, Mamba-based, and hybrid models. Each paradigm is defined by two inseparable components: the “task formulation” and the “computational architecture.” The task formulation specifies how the detection problem is transformed into a mathematical problem (e.g., bounding box regression, pixel-level segmentation, conditional denoising generation, etc.), while the computational architecture specifies which computational units are used to implement this formulation (e.g., CNN convolutions, self-attention mechanisms, state space models, etc.). For each paradigm, we analyze the potential conflicts between these paradigms and the physical characteristics of infrared weak targets, as well as the corresponding optimization pathways. Building upon this foundation, through a comprehensive survey of current algorithms, we have selected over 20 representative methods and established a cross-architecture performance comparison benchmark using five metrics (IoU, nIoU, Pd, Fa, and Params) across three datasets (NUAA-SIRST, NUDT-SIRST, and IRSTD-1k). Through three types of diagnostic experiments—performance benchmarking, evaluation of generalization capability across datasets, and robustness testing against synthetic interference sources—this study reveals cross-paradigm patterns, suggesting that within the settings and datasets examined, the primary performance bottleneck appears to have shifted from architectural complexity to the preservation of spatial information during the encoding stage and the learning of semantic discriminative representations. This review can serve as a valuable reference for both newcomers seeking to understand current trends and researchers exploring future directions for infrared weak target detection.