Objective The synergistic analysis of infrared and visible images holds significant theoretical and practical value for transcending the limitations of single-modal perception and constructing highly robust intelligent detection systems. Addressing the key challenges in visible-infrared image registration—namely, the scarcity of training data and significant modal disparities—the paper proposes a lightweight and efficient deep learning framework. The core contributions are threefold. First, we design a recurrent multiscale network architecture that substantially reduces the dependency on large-scale annotated data and improves adaptability across scales. Second, to bridge the feature representation gap between modalities, we enhance the high-frequency details of visible images via Discrete Cosine Transform (DCT) and extract robust features through a dual-path fusion mechanism. Finally, we introduce a sparse cross-attention mechanism based on Top-k selection, which focuses on critical regions while drastically reducing computational complexity. Extensive experiments demonstrate that the method achieves superior registration accuracy with limited data, alongside significantly enhanced robustness and generalization capability.

Methods To address these issues, a recurrent multiscale network framework for visible–infrared image registration was developed. The proposed framework contains three major components. First, a recurrent multiscale training architecture was designed to process image pairs iteratively across different resolutions. Through cyclic downsampling and repeated forward propagation, the model effectively expands training diversity while learning scale-invariant feature representations (Fig.1, Fig.2). This strategy reduces dependence on extensive datasets and improves adaptation to multiscale variations. Second, considering the feature distribution differences between visible and infrared images, frequency-domain enhancement was introduced using Discrete Cosine Transform (DCT). High-frequency information of visible images was selectively enhanced and reconstructed to improve edge and texture representation. A dual-path feature fusion mechanism combining Global Average Pooling (GAP) and Global Max Pooling (GMP) was further employed to jointly capture global contextual information and local salient responses. Third, a sparse cross-attention mechanism based on Top-k selection was introduced to focus on highly relevant regions while suppressing redundant information. By preserving only the most significant attention responses, computational complexity was reduced and matching robustness was enhanced.

Results and Discussions Experimental evaluation was conducted on four public visible–infrared datasets, including OSU, TNO, M3FD, and MSRS. Comparative experiments involved representative traditional approaches and recent deep learning-based registration methods. Qualitative results show that conventional approaches often produced sparse feature correspondences and exhibited noticeable mismatches under complex cross-modal scenarios. Recent learning-based methods improved matching density and robustness; however, the proposed framework generated more accurate and denser correspondences across different datasets, demonstrating superior registration capability under multimodal conditions (Fig.3). Quantitative evaluations were performed using Matching Precision (MP), Mean Error Distance (MED), and Root Mean Square Error (RMSE). Experimental results demonstrate that the proposed method consistently achieved the best overall performance across all datasets, yielding higher matching precision and lower geometric error metrics (Tab.1). The findings indicate improved registration accuracy and enhanced stability in cross-modal image alignment tasks. A sensitivity analysis was further performed to investigate the influence of different Top-k selection ratios (Tab.2). Experimental results indicate that excessively sparse attention maps suppressed useful matching information, whereas larger values introduced redundant responses and noise. The optimal balance between computational efficiency and registration performance was achieved when Top-k was set to 20%, resulting in the highest matching precision and the lowest geometric errors (Tab.2). Furthermore, ablation studies verified the effectiveness of each proposed component (Tab.3). Although frequency-domain enhancement provided moderate performance gains, the recurrent multiscale structure and sparse attention mechanism contributed more significantly to registration improvement. The complete framework achieved the best performance, confirming strong complementarity among different modules and validating the effectiveness of the overall design.

Conclusions A recurrent multiscale framework for visible–infrared image registration under limited training conditions was developed. By integrating multiscale cyclic learning, frequency-domain feature enhancement, and sparse attention mechanisms, the proposed method effectively reduces dependence on large datasets while improving feature representation and registration robustness. Experimental results verify superior performance in both matching precision and geometric error reduction across multiple datasets. The proposed framework demonstrates strong robustness, improved generalization capability, and high registration accuracy, providing an efficient solution for visible–infrared image alignment in practical applications.