RGB-IR object detection provides a promising solution for complex scenarios, such as remote sensing and low-light environments, by leveraging the complementary strengths of visible and infrared modalities. Despite significant advancements, two key challenges remain: (1) effectively integrating multi-modal features within lightweight frameworks to enable real-time performance and (2) fully utilizing multi-scale features, which are crucial for detecting objects of varying sizes but are often underexploited, leading to suboptimal accuracy. To address these challenges, we propose FQDNet, a novel RGB-IR object detection network that integrates an optimized fusion strategy with a Quad-Head detection framework. To enhance multi-modal feature fusion, we introduce a Channel Swap SCDown Block (CSSB) for initial feature interaction and a lightweight Spatial Channel Attention Fusion Module (SCAFM) to further refine the integration of complementary RGB-IR features. To improve multi-scale feature utilization, we designed the Dynamic-Weight-based Quad-Head Detector (DWQH), which dynamically assigns weights to different scales, enabling adaptive fusion and enhancing multi-scale feature representation. This mechanism significantly improves detection performance, particularly for small objects. Furthermore, to ensure real-time applicability, we incorporate lightweight optimizations, including the Partial Cross-Stage Pyramid (PCSP) and SCDown modules, which reduce computational complexity while maintaining high detection accuracy. FQDNet was evaluated on three public RGB-IR datasets—M3FD, VEDAI, and LLVIP—achieving mAP@[0.5:0.95] gains of 4.4%, 3.5%, and 3.1% over the baseline, with only a 0.4 M increase in parameters and 5.5 GFLOPs overhead. Compared to state-of-the-art RGB-IR object detection algorithms, our method strikes a better balance between detection accuracy and computational efficiency while exhibiting strong robustness across diverse detection scenarios.

In recent years, deep learning techniques have excelled in video action recognition. However, currently commonly used video action recognition models minimize the importance of different video frames and spatial regions within some specific frames when performing action recognition, which makes it difficult for the models to adequately extract spatiotemporal features from the video data. In this paper, an action recognition method based on improved residual convolutional neural networks (CNNs) for video frames and spatial attention modules is proposed to address this problem. The network can guide what and where to emphasize or suppress with essentially little computational cost using the video frame attention module and the spatial attention module. It also employs a two-level attention module to emphasize feature information along the temporal and spatial dimensions, respectively, highlighting the more important frames in the overall video sequence and the more important spatial regions in some specific frames. Specifically, we create the video frame and spatial attention map by successively adding the video frame attention module and the spatial attention module to aggregate the spatial and temporal dimensions of the intermediate feature maps of the CNNs to obtain different feature descriptors, thus directing the network to focus more on important video frames and more contributing spatial regions. The experimental results further show that the network performs well on the UCF-101 and HMDB-51 datasets.

Recent deep learning-based face recognition (FR) methods have demonstrated remarkable performance in high-resolution (HR) or down-sampled low-resolution (LR) tasks. However, these methods often exhibit disappointing speed-accuracy tradeoffs when deployed on real-world LR scenarios due to limited model generalization. To this end, we propose a lightweight Face Transformer framework for real-world LRFR named LLRFaceFormer. First, we propose a Transformers-as-Convolutions (TaC) network using a Transformer layer to replace the matrix multiplication in the standard convolutional process. This hybrid approach combines the strengths of Transformers and CNNs, allowing the TaC network to extract sufficient effective identity information via a global receptive field while adaptively discarding redundant homogeneous identity information on constructed LR faces. Second, we propose a Transformer-specific adaptive average procedure that incorporates tensor shape operations and a depthwise (DW) convolution. This procedure enables the LLRFaceFormer framework to focus on different regions of the input images. We also introduce an identity-aware simulator that generates real-world-like blurred LR scenarios during the training process to reduce the distribution discrepancy between the training LR faces and the testing real-world LR faces. The identity-aware simulator is simultaneously trained with the FR network with a cooperative training strategy. Further experiments illustrate the significantly superior speed-accuracy tradeoffs over existing LRFR methods with state-of-the-art (SOTA) performance on several LRFR benchmarks.

Infrared and visible image fusion (IVIF) is a key technique for synthesizing a comprehensive representation of a scene by exploiting diverse perceptual information. However, existing fusion methods encounter challenges in simultaneously preserving intricate texture details and extracting high-level semantic information–essential for downstream vision tasks. To address these issues, this article presents B2MFuse, a novel bi-branch multi-scale IVIF network based on joint semantics injection. The bi-branch consists of an interactive detail branch (ID-branch) and a parallel semantic branch (PS-branch), both featuring dual paths for infrared and visible modality. The former employs a channel exchange strategy that maximises the capture of modality-specific details while obtaining complementary features from the alternate modality. The latter efficiently captures semantic information and provides flexible scene knowledge guidance to the ID-branch, facilitating the subsequent top-tobottom multi-scale feature fusion and reconstruction process. A Spatial Weighted Channel Attention Fusion Module (SWCAFM) is then meticulously designed to enhance the integration of crucial fine-grained features across different scales. Furthermore, a scene-perception loss function is tailored to account for variations in the original image content. The synergy between B2MFuse’s advanced architecture and loss function ensures robust and superior fusion results in diverse environments, in particular enhancing human visual observation and supporting downstream visual tasks. Extensive evaluations on four public datasets demonstrate the superiority of our B2MFuse, compared to the state-of-the-art IVIF methods. The source code is available at https://github.com/arkymeng/B2MFuse.

Aerial object detection is a fundamental component of UAV-based surveillance and Earth observation, supporting applications such as traffic monitoring, urban management, and emergency search-and-rescue. However, complex backgrounds and pronounced scale variation in aerial imagery pose challenges to feature representation learning across shallow feature extraction, deep semantic modeling and multi-scale fusion. To tackle these challenges, we propose an Adaptive Dual-Refinement YOLO (ADR-YOLO), featuring two stage-specific refinements in the backbone and a scale-aware fusion strategy in the neck. At shallow stages, a Denoising Module (DNM) suppresses background-induced activations while preserving fine-grained structural cues through spatial structure modeling and element-wise modulation. In the deep backbone, a Gated Selective Attention (GSA) mechanism employs a contrast-enhanced attention formulation with semantic-adaptive gating to improve foreground discrimination under background interference. In the neck, a Channel-Compressed Fusion Module (CCFM) aligns multi-level features via bottleneck-style channel mapping, and leverages Global Response Normalization (GRN) to recalibrate channel responses and reduce fusion redundancy. Extensive experiments on VisDrone2019-DET, TinyPerson, and DIOR demonstrate that ADR-YOLO outperforms state-of-the-art detectors with marginal additional computational overhead. Our source code will be released on GitHub.