The research of infrared image small-target detection is of great significance to security monitoring, satellite remote sensing, infrared early warning, and precision guidance systems. However, small infrared targets occupy few pixels and lack color and texture features, which make the detection of small infrared targets extremely challenging. This paper proposes an effective single-stage infrared small-target detection method based on improved FCOS (Fully Convolutional One-Stage Object Detection) and spatio-temporal features. In view of the simple features of infrared small targets and the requirement of real-time detection, based on the standard FCOS network, we propose a lightweight network model combined with traditional filtering methods, whose response for small infrared targets is enhanced, and the background response is suppressed. At the same time, in order to eliminate the influence of static noise points in the infrared image on the detection of small infrared targets, time domain features are added to the improved FCOS network in the form of image sequences, so that the network can learn the spatio-temporal correlation features in the image sequence. Finally, compared with current typical infrared small-target detection methods, the comparative experiments show that the improved FCOS method proposed in this paper had better detection accuracy and real-time performance for infrared small targets.

Convolutional neural networks (CNNs) have made great achievements on computer vision tasks, especially the image classification. With the improvement of network structure and loss functions, the performance of image classification is getting higher and higher. The classic Softmax + cross-entropy loss has been the norm for training neural networks for years, which is calculated from the output probability of the ground-truth class. Then the network’s weight is updated by gradient calculation of the loss. However, after several epochs of training, the back-propagation errors usually become almost negligible. For the above considerations, we proposed that batch normalization with adjustable scale could be added after network output to alleviate the problem of vanishing gradient problem in deep learning. The experimental results show that our method can significantly improve the final classification accuracy on different network structures, and is also better than many other improved classification Loss.

Efficient infrared dim object detection has been challenged by low signal-to-noise ratios (SNRs). Traditional methods rely on the gradient difference and fixed-parameter model. These methods fail to adapt to sophisticated and variable situations in the real world. To tackle the issue, a deep learning method based on the spatio-temporal network is proposed in this paper. The model is established by the Convolutional Long Short-Term Memory cell (Conv-LSTM) and the 3D Convolution cell (3D-Conv). It is trained to learn the motion constraint of moving targets (spatio-temporal constraint module, called STM) and to fuse the multiscale local feature between the target and background (deep spatial features module, called DFM). In addition, a variable interval search module (state-aware module, called STAM) is added to the inference. The submodule decides to conduct a global search for images only if the target is lost due to fast motion, uncertain obstruction, and frame loss. Comprehensive experiments indicate that the proposed method achieves better performance over all baseline methods. On the mid-wave infrared datasets collected by the authors, the proposed method achieves a 95.87% detection rate. The SNR of the dataset is around 1–3 dB, and the background of the sequence includes sky, asphalt road, and buildings.

In response to problems such as large target scale variations, strong background noise, and blurred features leading by low contrast in infrared target detection in near space environments, this paper proposes an efficient detection model, YOLO-MARS, which is based on YOLOv8. The model introduces a Space-to-Depth (SPD) convolution module into the backbone section, which retains the detailed features of smaller targets by downsampling operations without information loss, alleviating the loss of the target feature caused by traditional downsampling. The Grouped Multi-Head Self-Attention (GMHSA) module is added after the backbone’s SPPF module to improve cross-scale global modeling capabilities for target area feature responses while suppressing complex thermal noise background interference. In addition, a Light Adaptive Spatial Feature Fusion (LASFF) detector head is designed to mitigate the scale sensitivity issue of infrared targets (especially smaller targets) in the feature pyramid. It uses a shared weighting mechanism to achieve adaptive fusion of multi-scale features, reducing computational complexity while improving target localization and classification accuracy. To address the extreme scarcity of near space data, we integrated 284 near space images with the HIT-UAV dataset through physical equivalence analysis (atmospheric transmittance, contrast, and signal-to-noise ratio) to construct the NS-HIT dataset. The experimental results show that ${mAP}_{@0.5}$ increases by 5.4% and the number of parameters only increase 10% using YOLO-MARS compared to YOLOv8. YOLO-MARS improves the accuracy of detection significantly while considering the requirements of model complexity, which provides an efficient and reliable solution for applications in near space infrared target detection.

Infrared small target detection is a critical task in remote sensing. However, it remains highly challenging due to low contrast, heavy background clutter, and large variations in target scale. Traditional convolutional networks are inadequate for joint modeling, as they cannot effectively capture both fine structural details and global contextual dependencies. To address these issues, we propose FSGPNet, a frequency–spatial domain jointly guided perceptual network that explicitly exploits complementary representations in both the frequency and spatial domains. Specifically, a Frequency–Spatial Enhancement Module (FSEM) is introduced to strengthen target details while suppressing background interference through high-frequency enhancement and Perona–Malik diffusion. To enhance global context modeling, we propose a Multi-Scale Global Perception (MSGP) module that integrates non-local attention with multi-scale dilated convolutions, enabling robust background modeling. Furthermore, a Gabor Transformer Attention Module (GTAM) is designed to achieve selective frequency–spatial feature aggregation via self-attention over multi-directional and multi-scale Gabor responses, effectively highlighting discriminative structures of various small targets. Extensive experiments are conducted on two benchmark datasets (IRSTD-1K and NUDT-SIRST) that cover typical remote sensing infrared scenarios. Quantitative and qualitative results demonstrate that FSGPNet consistently outperforms state-of-the-art methods across multiple evaluation metrics. These findings validate the effectiveness and robustness of the proposed FSGPNet for detecting small infrared targets in remote sensing applications.

Infrared small target detection is a critical task in remote sensing, where targets typically span only a few pixels and are easily submerged by complex background clutter. Traditional convolutional networks predominantly rely on data-driven feature learning or computationally expensive global attention mechanisms. In this paper, we propose the Anomaly-Driven Gated Fusion Network (ADGFNet), a lightweight end-to-end framework that explicitly embeds the statistical characteristics into the feature extraction and fusion pipeline. The core of ADGFNet comprises two synergistic modules: Local Anomaly Block (LAB), which decomposes anomaly perception into a Local Contrast Branch capturing multi-scale intensity saliency and a Gradient Prior Branch extracting boundary discontinuities via Sobel operators; and Anomaly-Guided Feature Pyramid Neck (AG-FPN), which reuses the spatial anomaly maps generated by the LAB as guidance gates to perform spatially selective cross-scale feature fusion without requiring additional learnable spatial-attention modules. Comprehensive evaluations on the NUDT-SIRST and IRSTD-1K benchmarks demonstrate that ADGFNet achieves a superior accuracy–efficiency trade-off, attaining the best nIoU and the lowest false alarm rates with only 0.58 M parameters. With TensorRT FP16 acceleration, ADGFNet and its ultra-lightweight variant, ADGFNet-Lite (0.12 M parameters), enable real-time inference without appreciable loss of detection accuracy on an NVIDIA Jetson AGX Xavier.