Infrared small target detection (IRSTD) has progressed significantly in spatial-domain learning. However, single-frame images’ limited spatial semantics impair discrimination between targets and similar noise while complicating integrity detection of large-scale target. To address these, we propose Global Spatial-Frequency Attention Network (GSFANet), which enhances the distribution difference between targets and noise from a frequency-domain perspective while preserving spatial information integrity. The core innovations consist of three modules: 1) Parametric Wavelet Downsampling (PWD), preserving small target details during frequency refinement to prevent feature fragmentation; 2) Hierarchical Gated Kernel Attention (HGKA), capturing cross-level frequency relationships through Cross-channel Kernel Attention (C²K) and maintaining spatial coherence via Cross-spatial Gate Attention (CSG), effectively bridging semantic gaps across layers; 3) Adaptive Frequency-Decoupled Fusion (AdaFD), dynamically fusing target-associated frequency components while suppressing noise. We further develop AdaFL Loss to balance multi-scale target gradients and stabilize training. Experiments on three benchmark datasets demonstrate GSFANet’s superior detection performance and enhanced segmentation robustness in complex scenarios compared to state-of-the-art methods. Our code will be made public at https://github.com/dengfa02/GSFANet_IRSTD.

Object detection is of wide application for its capability in recognizing and locating targets in scenes. Under heavy hazy conditions, however, the detection performance on RGB images will greatly degrade since the image contents are polluted. In this work, we propose an imaging system to acquire three-channel images in the shortwave infrared (SWIR) spectrum to facilitate object detection under hazy conditions. The system captures SWIR images in the form of pseudo color that are most suitable for detecting objects, such as pedestrians and vehicles. Two different types of filters, i.e., liquid crystal tunable filter (LCTF) and optical filters, are employed in our imaging system design. We use the LCTF to acquire narrowband hyperspectral images, which are fed into a band simulation model to generate wideband images for optimal band selection. We present a specific measure called recognition and localization (RL) score to evaluate the detection performance of three-band combinations. Based on the measure, optimal bands are determined using an efficient searching algorithm. Then, we customize three optical filters and install them on a filter wheel, with which we can acquire three-channel images in the SWIR spectrum. The effectiveness of our imaging system is evaluated on a self-collected RGB-SWIR image dataset. Experimental results indicate that, compared with the RGB images after haze removal, the three-channel SWIR images acquired by our imaging system are of higher quality and can achieve better object detection performance under hazy conditions.

Motion Infrared Small Target Detection (MIRSTD) leverages multi-frame temporal dependencies to improve detection robustness. However, existing methods have difficulty modeling global consistency and achieving precise alignment in complex motion and large displacement scenarios, leading to dispersed target representations and higher error rates. To address these challenges, we propose Dynamic Query Aligner (DQAligner), which introduces global random large-displacement augmentation and a cross-scale bidirectional shared attention mechanism to enhance inter-frame consistency. A dynamic receptive field pyramid deformable convolution decomposes complex multi-scale motions, enabling precise target alignment. Furthermore, class query memory serves as the generalized residual form of deformable convolution, which iteratively learns dynamic query representations to facilitate global target localization within each frame and maintain semantic consistency across frames. DQAligner achieves a paradigm shift from rigid alignment to flexible matching, and significantly boosts detection performance in large displacement and dynamic scenarios. Experiments on extensive stationary and moving platform datasets show that DQAligner outperforms existing methods, especially under complex motion and low signal-noise-rate conditions. Code will be available at https://github.com/dengfa02/DQAligner_MIRSTD.

Hyperspectral image (HSI) classification is a critical task in remote sensing. Recently, deep learning (DL) methods, particularly state space models (SSMs) like Mamba, have garnered significant attention due to their linear computational complexity in handling long sequences. However, processing the rich yet redundant spectral bands in HSIs remains challenging for these approaches, often resulting in models with excessive parameters. Moreover, existing DL-based methods frequently overlook explicit cross-channel dependencies. This limitation hinders their ability to effectively aggregate relevant information from redundant spectral bands. Additionally, the heavy reliance of DL-based models on extensive training data conflicts with the scarcity of labeled samples in HSI scenarios. This necessitates the incorporation of prior knowledge to mitigate data limitations. To address these challenges, we propose SliMamba, a lightweight Convolution-Mamba architecture for HSI classification (HSIC). SliMamba introduces two key components: spectral selection convolution (SSC) and overlapped-centering mamba (OC-Mamba). SSC swaps dimensions between spectral and spatial axes while preserving the original spectral information. It then utilizes a small set of convolutional kernels to learn spectral weights within the receptive field, thereby achieving cross-channel feature representation. OC-Mamba integrates prior knowledge of the center pixel with local feature enhancement. It combines one-pixel overlapped SSM (O-SSM) and centering SSM (C-SSM) to extract global–local and spectral–spatial features from HSI feature maps. Extensive evaluations on four benchmark datasets demonstrate that SliMamba consistently outperforms state-of-the-art methods in classification accuracy while significantly reducing model complexity. The source code is available at https://github.com/flyzzie/SliMamba

The Automatic Identification System (AIS) extends traditional measurement methods, enabling real-time situational awareness at intersections. However, most methods rely on extensive historical data from multiple intersections, which conflicts with the limited sensing range of onboard measurement. Moreover, state-of-the-art (SOTA) models suffer from error accumulation, high prediction uncertainty, and inadequate global variate modeling, limiting trend awareness and generalization across measurement scales. To overcome these, we propose iTentformer, an encoder-only model focusing on short-term vessel behavior at intersection waterways. Key innovations include: (1) a variate-centered encoder to capture distribution differences before intersections, and (2) integration of the predicted course as a trend prior, highlighting pattern changes before intersections. These components are optimized through variate representation fusion and multi-task learning. Extensive experiments show that iTentformer reduces ADE by 35% and FDE by 30% compared to SOTA Transformer-based models, improving trajectory abstraction and reducing estimation uncertainty through strong trend awareness. Notably, vessel navigational status at intersections significantly impacts measurement distribution. Additionally, its improved RUL prediction highlights its potential for broader measurement applications. The code is available at github.com/dengfa02/iTentformer.