27 October 2020 in Optics Letters
Laser damage in fused silica, particularly ultraviolet laser damage, is still a key problem limiting the development of high-power laser systems. In this Letter, a combined process of chemical etching and laser polishing was applied to ground fused silica. A super-smooth surface with a root-mean-square roughness of 0.25 nm was achieved through this combined process. Furthermore, the combined process can reduce the introduction of photoactive metal impurity elements, destructive defects, and chemical-structure defects, resulting in a 0% probability damage threshold nearly 33% higher than a conventional chemical mechanical polished sample for a 7.6 ns pulse at a wavelength of 355 nm.
14 February 2025 in ISPRS Journal of Photogrammetry and Remote Sensing
4 September 2024 in International Journal of Applied Earth Observation and Geoinformation
1 January 2026 in IEEE Transactions on Instrumentation and Measurement
4 February 2026 in SSRN Electronic Journal
Unsupervised Video Anomaly Detection (UVAD), a critical technology in surveillance and security, aims to automatically identify abnormal events in videos. Common approaches include visual modeling (e.g., reconstruction and prediction), normality modeling (e.g., probabilistic and distance-based models), and specific self-supervised proxy tasks. However, these self-supervised methods often struggle to learn rich video representations directly and typically rely on negative samples or additional components (e.g., optical flow, object detection, or skeleton tracking). To address these limitations, we propose SimAdaM—a simple yet efficient representation learning framework that integrates non-contrastive learning with adaptive spatio-temporal masked modeling. Specifically, we introduce a motion-aware masking strategy that dynamically adjusts masking ratios based on gradients, thereby strengthening the information bottleneck during training and forcing the model to learn more informative representations. For the non-contrastive learning component, we employ an asymmetric siamese network with a stop-gradient strategy, enabling effective dispersion of normal sample features without relying on negative examples. Built upon a strong baseline, our enhanced feature extraction strategy demonstrates significant improvements. Extensive experiments validate the effectiveness and robustness of the proposed approach, outperforming existing frame-level methods by 1\%–3\% and achieving micro-AUC scores of 84.7\% and 93.0\% on ShanghaiTech and Avenue datasets, respectively, while maintaining the same computational efficiency as the baseline during inference. The code is to be released publicly.