10 February 2025 in IEEE Access
Industrial anomaly detection plays a critical role in computer vision, driving advancements across various sectors. In recent works in this field, there are trade-offs between computational complexity and anomaly detection accuracy for effective deployment. To achieve both reduced latency and improved accuracy, we propose a simple reconstruction network that learns local clusters centered at core-features derived from PatchCore. By leveraging these features that represent local data distributions and their adjacent patch-features, our network effectively enhances anomaly detection along with reduced inference time. Specifically, the proposed reconstruction network is designed to reconstruct core-features from their neighbors, effectively acting as a denoiser for Neighbor Noise. So neighbor noise is defined as the distance between a core-feature and its adjacent patch-features, which can be regarded as a noise added to core-faeture. By reducing this noise, our network creates tightly clustered reconstruction of core-features and their neighbors, excluding reconstructions of anomaly features. Evaluated on MVTecAD and MPDD datasets, our approach outperforms existing methods in both inference efficiency and accuracy. The code will be made available upon acceptance.
11 March 2025 in IEEE Access
Face anti-spoofing (FAS) techniques are crucial for safeguarding face recognition systems against an ever-evolving landscape of spoofing attacks. While existing methods have made strides in detecting known attacks, robustness against unknown, potentially more sophisticated attack types remains a critical and unresolved challenge in the field. In this paper, we propose a novel framework integrating token-wise asymmetric contrastive learning with angular margin loss to enhance the robustness of FAS models against unknown attack types. The key idea to resolve the problem is to learn a feature space where live face features are densely distributed, whereas spoof face features are more dispersed. This is achieved through two novel strategies: (1) asymmetric contrastive learning, encouraging the FAS model to learn a compact distribution for live face features while relaxing the constraint on the distribution of spoof face features, and (2) token-wise learning, focusing on capturing intrinsic liveness cues from local region rather than identity or facial-related features. Additionally, an angular margin loss is incorporated to enhance the discriminative power of the learned features. Extensive experiments on public benchmark datasets demonstrate the superiority of our FAS model over state-of-the-art methods in cross-attack scenarios, showcasing its strong robustness to unknown attacks while maintaining unseen domain generalization capability.
20 May 2026 in IEEE Access
Image steganalysis, the detection of hidden information embedded in digital images, is a core component of modern cybersecurity and digital forensics. Recent residual Transformer architectures, such as Pixel-Difference-Convolution and Enhanced-Transformer-Network (PENet) [1], achieve strong detection accuracy, but their computational and memory demands hinder deployment in real-time or resource-constrained settings. In this paper, we present PENet+, a lightweight steganalysis framework that preserves PENet’s discriminative structure while substantially improving efficiency. Instead of redesigning or compressing the attention blocks, we retain PENet’s self-attention topology for reproducibility and introduce a classifier-streamlining stage that progressively narrows the SPP→FC1 input channels, yielding large reductions in parameters and FLOPs with negligible accuracy loss. Here, SPP denotes spatial pyramid pooling, FC1 denotes the first fully connected layer after SPP, and HPF denotes high pass filter. We further refine the HPF stem with an activation-aware selection mechanism that aggregates high-pass filter responses in an early stage and selects a balanced SRM–Gabor top-K subset, and we replace PENet’s backbone with a MobileNetV2-style inverted residual network for efficient spatial processing. A balanced HPF configuration with K=31 filters (16 Gabor + 15 SRM) matches or surpasses the accuracy of heavier configurations while lowering compute. Finally, we revisit activation design and motivate PReLU from a steganalysis standpoint, arguing that preserving negative responses is beneficial for capturing weak stego cues that ReLU would otherwise suppress. On the ALASKA2 [2] QF90 color benchmark at 512×512 resolution, PENet+ achieves up to a 45.5% reduction in parameters and approximately 97% fewer FLOPs while maintaining accuracy comparable to PENet. These results indicate that the proposed architecture offers a practical potential toward real-time, on-device steganalysis.
12 January 2026 in IEEE Access
Diffusion models achieve state-of-the-art performance in image denoising but are hindered by a slow, iterative sampling process. While advanced, training-free samplers and optimized schedules offer significant acceleration, their validation has been focused on image generation. Their applicability to pre-trained denoising models, which operate on conditional inputs, remains a critical, unverified gap. We address this gap by applying and validating a post-hoc timestep optimization that accelerates pre-trained denoising models without retraining. By minimizing the discretization error of the underlying diffusion ODE, this method finds a non-uniform optimal schedule for a given number of function evaluations (NFE). Our extensive experiments on diverse models for real-world, weather, and Gaussian noise demonstrate that the proposed optimized schedule effectively reduces the required NFE. We demonstrate that it is possible to reduce the number of sampling steps while significantly preserving, and often improving, restoration quality as measured by PSNR, SSIM, and LPIPS. This work presents and validates a practical, training-free strategy for substantially accelerating diffusion-based denoising, enhancing its viability for real-world applications. Our code is available at: https://github.com/cv-cho/Accelerating-Diffusion-Denoising-Models.
31 October 2025 in IEEE Access
Out-of-Distribution (OOD) detection is critical for ensuring the reliability of semantic segmentation models in safety-critical autonomous driving scenarios. Despite recent advances, existing state-of-the-art OOD segmentation methods fundamentally rely on local features and suffer from a critical lack of contextual understanding in complex road environments. A representative example is distant scenes with small objects that require contextual reasoning to distinguish them from background elements. To evaluate such complex and challenging cases, we construct a dedicated subset for robustness assessment. The root cause stems from their inability to perform contextual semantic reasoning about object appropriateness in road environments. To address these fundamental limitations, we propose a novel multi-agent visual reasoning framework that leverages the powerful contextual understanding and semantic reasoning capabilities of Vision-Language Models (VLMs). Our framework decomposes the OOD detection task into specialized subtasks handled by multiple expert agents. This approach fundamentally shifts from local pattern recognition to in-context understanding-based OOD detection, enabling the system to understand not just what is anomalous, but why it is inappropriate for the given road context. Extensive experiments demonstrate that our framework significantly outperforms existing methods, particularly in challenging scenarios, while providing interpretable reasoning for safety-critical applications.
2 May 2026 in SSRN Electronic Journal
Effective code generation depends not only on model capability but also on how problems are represented at inference time. Existing methods improve performance by adding reasoning steps or increasing sample diversity, yet they often leave the task description itself largely unchanged. We study code generation through structured task reformulation, where a problem is rewritten into a representation that preserves semantic content while making its algorithmic structure easier for the solver to recognize and implement. As a concrete instance, we build on narrative-based reformulation, which reorganizes fragmented statements into coherent task overviews, constraints, and example input-output descriptions. Across 11 language models and three execution-based benchmarks, HumanEval, LiveCodeBench, and CodeForces, structured reformulation consistently improves pass@10 over repeated sampling and reasoning-oriented baselines. Further analyses show that the gains arise through two complementary effects: better agreement with correct algorithmic strategies and fewer implementation-level errors once an appropriate strategy is selected. The results also show that performance depends on coherence across representation components, alignment with the solver’s reasoning space, and structural reorganization rather than narrative style alone.