논문

Digital screen photography frequently suffers from moiré patterns caused by frequency aliasing between camera sensors and display pixels. Removing moiré artifacts is challenging due to their varying color, shape, and scale. Existing methods primarily operate in the spatial domain and thus overlook crucial frequency-domain information inherent in moiré patterns. To address this, we propose Spectrally Empowered Demoiréing Transformers (SEDT), a U-shaped Transformer integrating a Frequency Domain Modulator (FDM) within each Transformer block. The FDM performs spectral filtering through learnable modulation of frequency components, such as channel-wise affine transformations and band-wise frequency masking, enabling effective extraction of moiré-related features across multiple scales and frequency ranges. Evaluations on the TIP-2018 dataset demonstrate that SEDT significantly outperforms existing state-of-the-art models in both PSNR and SSIM, highlighting the potential of frequency-domain enhancements in Transformer architectures for advanced image restoration tasks. The code is available at https://github.com/sylee-skhu/SEDT.

Robust fall detection is critical in safety-sensitive contexts such as elderly care and industrial environments. Recent fall detection methods leverage 2D human pose estimation, which encodes posture as compact skeletal keypoints. However, these systems are vulnerable to performance degradation when body parts are occluded or keypoints are missing, often due to environmental constraints or inaccuracies in pose estimation. To address this, we propose a fall detection framework based on single-frame 2D human pose estimation and a skeleton-aware graph attention network (StoneGAT). StoneGAT enhances conventional graph attention network (GAT) by incorporating edge features based on bone-related information, such as bone lengths, joint angles, and confidence-based metrics. In addition, we introduce a training strategy named PointOut, which probabilistically drops node features during training to encourage structure-aware learning, thereby improving model robustness. Experiments on a combined dataset of AI-hub and in-house dataset demonstrate that StoneGAT with PointOut outperforms baseline models, including standard MLP, graph convolutional network (GCN) and GAT, especially under severe occlusion. Ablation studies confirm the effectiveness of edge attributes and the PointOut strategy.

Tiny object detection (TOD) plays a vital role in robotics and control systems, particularly in applications such as dronebased surveillance, autonomous navigation, and intelligent monitoring. Detecting extremely small objects in aerial imagery poses significant challenges due to limited spatial detail and insufficient contextual information. While feature pyramid networks (FPNs) have proven effective in handling multi-scale features, conventional FPN-based approaches often suffer from reduced spatial precision and semantic richness, factors critical for reliable tiny object detection. To address these limitations, we propose an enhanced FPN architecture that integrates deformable convolution (DeformConv) and vision state space modules (VSSM). DeformConv replaces standard fixed-kernel convolutions in the lateral connections, enabling adaptive focus on sparse spatial locations to better capture geometric variations and fine-grained spatial features. In parallel, VSSM modules are incorporated into the top-down pathway to model long-range contextual dependencies, thereby mitigating the semantic information loss typically encountered in traditional FPNs. Experiments conducted on the VisDrone2019 and AI-TOD v2 datasets demonstrate that the proposed architecture consistently outperforms conventional FPN-based detectors, validating the robustness and effectiveness of our feature refinement strategy for tiny object detection.

This study proposes a campus tour mobile application called DeepCampus that leverages deep learning-based vision technologies to enhance user engagement and interaction. The proposed application introduces two core features: a photo mission, where users take photos at designated campus locations, and a commemorative photo creation function that transforms these photos into character-based styles. The photo mission relies on a CNN-based place recognition model, trained on a custom-built campus dataset, to verify mission completion. The commemorative photo creation function uses a diffusion model to generate personalized, stylized photos, providing users with unique records of their campus experience. By integrating place recognition and generative models, the application encourages active exploration. The effectiveness of the proposed system is validated through experiments on the custom dataset, demonstrating its technical reliability and potential to enhance campus tours.

Transformer and Mamba-based models have shown promising results in field of video understanding. However, their excessive complexity often leads to the problem of computational overhead and memory explosion. To address this issue, we propose a novel model called Mamba VisionTSM, which integrates the MambaVision architecture with the Temporal Shift Module (TSM). TSM has been effectively used to enable temporal modeling in 2D CNN models. In this work, we extend its application to the hybrid Mamba Vision model, allowing it to efficiently process temporal information while retaining the long-range spatial dependency capabilities of Mamba Vision. Through experimental results on the Something-Something V2 dataset, we demonstrate that our approach achieves comparable or superior performance compared to existing Transformer-based approaches in action recognition, highlighting its potential for broader video understanding tasks.

In this study, we present the Discriminative Temporal Shift Module (D-TSM), an enhancement of the Temporal Shift Module (TSM) for action recognition. TSM has limitations in capturing intricate temporal dynamics due to its simplistic feature shifting. D-TSM addresses this by introducing a subtraction operation before the shifting. This enables the extraction of discriminative features between adjacent frames, thereby allowing for effective action recognition where subtle motions serve as crucial cues. It preserves TSM???foundational philosophy, prioritizing minimal computational overhead and no additional parameters. Our experiments demonstrate that D-TSM significantly improves performance of TSM and outperforms other leading 2D CNN-based methods.

Action recognition is one of the representative perception tasks for robot application, but it still remains challenging due to complex temporal dynamics. Although temporal shift module (TSM) has been considered to be one of the best 2D CNN based architecture for temporal modeling, its inherent structural simplicity limits performance and has room for improvement. To mitigate this issue while following TSM's philosophy, this paper presents a variant of TSM, termed as Discriminative TSM (D-TSM), with a focus on capturing dis-criminative features for motion pattern. Going further from the naive shift operation in TSM, our D-TSM explicitly transforms shifted features by applying element-wise subtraction. This simple approach is effective to create discriminative features between adjacent frames with a small extra computational cost and zero parameter. The experiments on Something-Something and Jester datasets demonstrate that our D-TSM outperforms TSM and achieves competitive performance with low FLOPs against other methods.

Siamese convolutional neural networks (SCNNs) has been considered as among the best deep learning architectures for visual object verification. However, these models involve the drawback that each branch extracts features independently without considering the other branch, which sometimes lead to unsatisfactory performance. In this study, we propose a new architecture called an extended SCNN (ESCNN) that addresses this limitation by learning both independent and relative features for a pair of images. ESCNNs also have a feature augmentation architecture that exploits the multi-level features of the underlying SCNN. The results of feature visualization showed that the proposed ESCNN can encode relative and discriminative information for the two input images at multi-level scales. Finally, we applied an ESCNN model to a person verification problem, and the experimental results indicate that the ESCNN achived an accuracy of 97.7%, which outperformed an SCNN model with 91.4% accuracy. The results of ablation studies also showed that a small version of the ESCNN performed 5.6% better than an SCNN model.

When multiple object tracking (MOT) based on the tracking-by-detection paradigm is implemented, the similarity metric between the current detections and existing tracks plays an essential role. Most of the MOT schemes based on a deep neural network learn the similarity metric using a Siamese architecture, but the plain Siamese architecture might not be enough owing to its structural simplicity and lack of motion information. This paper aims to propose a new MOT scheme to overcome the existing problems in the conventional MOTs. Feature pyramid Siamese network (FPSN) is proposed to address the structural simplicity. The FPSN is inspired by a feature pyramid network (FPN) and it extends the Siamese network by applying FPN to the plain Siamese architecture and by developing a new multi-level discriminative feature. A spatiotemporal motion feature is added to the FPSN to overcome the lack of motion information and to enhance the performance in MOT. Thus, FPSN-MOT considers not only the appearance feature but also motion information. Finally, FPSN-MOT is applied to the public MOT challenge benchmark problems and its performance is compared to that of the other state-of-the-art MOT methods.

This study focuses on the problem of robust adaptive synchronization of uncertain bilinear chaotic systems. A Takagi?諛챬geno fuzzy bilinear system (TSFBS) is employed herein to describe a bilinear chaotic system. A robust adaptive observer, which estimates the states of the TSFBS, is also developed. Advanced adaptive laws using a projection operator are designed to achieve both the robustness for the external disturbances and the adaptation of unknown system parameters. A comparison with the existing observer shows that the proposed observer can achieve a faster parameter adaptation and a robust synchronization for an uncertain TSFBS with disturbances when the adaptive laws are utilized. The asymptotic stability and the robust performance of the error dynamics are guaranteed by some assumptions and the Lyapunov stability theory. We verify the effectiveness of the proposed scheme using examples of the generalized Lorenz system in various aspects.