This paper proposes an efficient algorithm for real-time multi-modal image matching based on a lightweight feature fusion network, targeting the challenges of multi-modal image matching in multi-source data analysis. The algorithm addresses significant multi-modal feature differences and real-time processing limitations by incorporating key technologies including reparameterization in convolutional neural networks, multi-scale image pyramids, and feature fusion modules. The matching process employs a coarse-to-fine strategy, ensuring robust performance in complex environments. Experimental results using multi-modal datasets demonstrate that the proposed algorithm achieves superior accuracy and speed, with a success rate of 98.3% and an average matching time of 30.51 ms per 500×500 image pair. These results highlight the practical value and strong generalization capability of the algorithm in real-time applications.

Few-shot learning enables rapid recognition of novel categories with limited samples, mimicking human cognition. However, existing unimodal few-shot image recognition methods often fail to maintain robustness under complex maritime disturbances. To address these challenges, the dual-stream feature map reconstruction network (DFRN) is proposed to address few-shot visible-infrared ship recognition. In this model, visible and infrared features are projected into a shared subspace for common semantics and into extended subspaces to extract modality-specific details. These parallel streams are then leveraged for collaborative reconstruction and fusion. Additionally, a full prototype constraint module (FPCM) is introduced to enhance intra-modality feature separability. A bidirectional probability collaborative learning module (BPCLM) is employed to facilitate mutual knowledge transfer across modalities. Furthermore, YTShip-10K is constructed: a large-scale ship dataset comprising 10083 paired visible and infrared images covering 101 fine-grained categories across five maritime scenarios. Experiments show that our model achieves 81.20% and 95.36% accuracy on 1-shot and 5-shot tasks, outperforming unimodal baselines by 4.94% and 5.29%, respectively. The YTShip-10K dataset and code will be publicly released.

Bird’s-eye-view (BEV) representations have become a widely adopted paradigm for light detection and ranging (LiDAR)-based 3D object detection due to their efficiency and structured spatial layout. However, most existing BEV-based detectors rely on convolutional neural networks (CNNs) for BEV feature extraction and fusion, which primarily capture local spatial contexts and may be limited in exploiting global contextual information in complex scenes. To address this limitation, a method is proposed in which a lightweight BEV transformer (LBT) is integrated into the BEV feature learning process to enhance global context modeling capability. The proposed LBT follows a plug-and-play design and can be easily integrated into existing BEV-based detectors. The proposed method is implemented on the CenterPoint framework and is evaluated on a standard LiDAR-based 3D object detection benchmark. Experimental results demonstrate performance improvements over the CNN-only baseline, indicating that incorporating lightweight global context modeling in the BEV space is an effective and practical way to enhance LiDAR-based 3D object detection.

In light of the limitations of single-source remote sensing data for accurate Earth observation, integrating hyperspectral image (HSI) and synthetic aperture radar (SAR) data has been shown to be meaningful for land-cover classification task. Nevertheless, owing to the substantial disparities in imaging mechanisms and data attributes across different sources, existing classification methods still encounter certain challenges in adaptability and collaboration for frequency feature extraction and heterogeneous information fusion. As such, a new progressive frequency-division network is presented for joint classification of HSI and SAR data. Firstly, a progressive frequency decomposition module is designed to fully explore high- and low-frequency information of multisource remote sensing data, and minimize redundant information, thus achieving complementary and discriminative fusion feature extraction. After that, with an effective bidirectional propagation mechanism, a multiscale interaction module is further built to collaborate local and global information and explore more discriminative multiscale representations, thereby more accurately representing and classifying complex land covers. Experimental results for two public datasets containing HSI and SAR data reveal that the proposed model outperforms other competitors.

Based on analyzing the storage structure of polarimetric interferometry synthetic aperture radar (Pol-InSAR) data, this study proposed a multi-core parallel Pol-InSAR data processing method (MCP-PIDPM). In order to improve the processing speed of Pol-InSAR data, this paper presented a multi-core parallel Pol-InSAR data processing scheme, and constructed its processing framework in the research. The key technology of the multi-core parallel Pol-InSAR data processing was discussed, a buffer detection splitting method for Pol-InSAR image was proposed, and improved the self-adaptive phase unwrapping method. At last, a multi-core parallel Pol-InSAR data processing system (MCP-PIDPS) was developed, and the proposed MCP-PIDPM method was tested and validated through experiments. The system based on the above method assigns tasks and loads reasonably to each processor and uses space to exchange time. It is proved by numerical experiments that the efficiency of Pol-InSAR data processing is increased by more than 10 times. The research breaks through the bottleneck which restricts the efficient application of Pol-InSAR, and provides a feasible solution for the efficient processing of Pol-InSAR data.

In the realms of computer vision and remote sensing, the matching of images and point clouds poses significant challenges due to modality discrepancies. This study introduces a cross-modal consistency network, detector-free image and point cloud matching via diffusion-guided cross-modal consistency, 2D3D-DiffMatch, leveraging diffusion prior information to enhance feature extraction consistency and alignment across modalities. To strengthen cross-modal consistency in complex scenes, diffusion priors generated by a pre-trained diffusion model are used to guide the feature extraction process toward semantically and geometrically consistent representations. These representations are further refined through a hierarchical fusion process, in which the most consistent diffusion features are adaptively selected using centered kernel alignment(CKA) and integrated with multi-scale backbone features, thereby mitigating the impact of modality gaps. Furthermore, to address feature-space misalignment between images and point clouds, we propose a cross-modal feature consistency loss that adaptively constrains correspondences, separates positive and negative pairs, and optimizes the agreement of positive pairs, enabling high-quality, detector-free matching. Experimental results on the 7Scenes and RGB-D Scenes V2 Datasets demonstrate superior registration recall rates of 81.2% and 61.0%, respectively, outperforming state-of-the-art methods and exhibiting robustness in challenging scenarios.This research advances the collaborative processing of multi-modal data, offering a robust solution for image–point cloud matching in challenging scenarios.

Unsupervised multimodal remote sensing change detection is challenging due to severe cross-modality discrepancies and unreliable pixel-wise correspondence. Existing methods usually treat cross-modal alignment and change discrimination as separate processes, which may lead to error accumulation and limited robustness under strong modality differences. In this paper, we propose a unified unsupervised framework that tightly couples structural alignment and change discrimination. A non-local graph alignment (NLG) module is introduced to establish structure-preserving cross-modal correspondence by modeling non-local spatial relationships. Meanwhile, a global-local state (GLS) discrimination module based on state-space modeling is designed to capture both long-range dependency patterns and fine-grained local variations of changes. The two modules are iteratively optimized in an end-to-end manner, eliminating the reliance on pseudo-label generation and explicit feature differencing. Extensive experiments on multimodal benchmarks demonstrate that the proposed method consistently outperforms state-of-the-art unsupervised multimodal change detection approaches.

Remote sensing change detection (RSCD) plays a critical role in disaster assessment, land use monitoring, and environmental analysis. Despite notable progress with deep learning, especially convolutional neural networks (CNNS) and transformer-based models, existing approaches still face challenges such as false detections in multi-scale objects and cross-scale semantic inconsistencies. To address these limitations, we propose DMF-CDNet, a dual-stream multi-scale fusion change detection built on a resnet-18 backbone. The model integrates two key modules: the feature-enhanced spatial-spectral feature coordination (FE-SSFC) module, which combines pyramid split attention module with residuals(PSAR)-based multi-scale convolution and statistically guided enhancement to highlight true changes and suppress noise, and the dual-branch decoding module (DDM), which incorporates guidance and channel rearrangement strategies to improve semantic consistency and boundary preservation. Through progressive decoding and cross-scale fusion, the network achieves more accurate localization of change regions and finer boundary recovery. Experimental results on the LEVIR-CD, HRCUS-CD, and SYSU-CD datasets demonstrate that DMF-CDNet achieves F1-scores of 91.52%, 74.92%, and 81.96%, respectively, confirming its effectiveness in multi-scale modeling and fine-grained change detection for complex RSCD scenarios.

Face detection in distributed edge computing is confronting challenges such as loss of facial features due to occlusions, insufficient feature of small-scale faces, and difficulty in efficient deployment on distributed terminals. In this paper, an improved you only look once version 12 nano (YOLOv12n) is proposed, specifically designed to promote the detection of occluded and small-scale faces in distributed edge computing. A spatial-to-depth convolution (SPD-Conv) is introduced to optimize feature extraction and reconstruct feature encoding. A cross stage partial receptive field enhancement module (C3RFEM) is integrated to strengthen modeling capability for occluded and small-scale faces by multi-branch dilated convolutions. A multi-patch Monte Carlo attention (MPMCA) is proposed to elevate robustness and adaptability by random perturbations. A shape-aware normalized Wasserstein distance (SA-NWD) is adopted to increase the accuracy of localization and shape modeling for small targets. Experimental results demonstrate the proposed model achieves improvements in mean average precision 50 (mAP50) with increase of 0.92%, 1.20% and 3.44% on the Easy, Medium and Hard subsets of the WiderFace dataset compared to the original YOLOv12n, with enhanced robustness and adaptability in distributed edge computing.