| 編號： | 2 |
| 期刊或研討會名稱： |
The Seventh International Symposium on Computer, Consumer and Control
台灣
|
| 著作名稱： | Enhancing Micro-Expression Recognition through Optimized Apex Frame Spotting: A Pair-wise Difference Maximization Approach
|
| 年度： | 2025 |
| 類別： |
會議論文
|
| 摘要： | Micro-expressions are subtle facial movements characterized by their extremely brief duration (approximately 0.2-0.5 seconds) and resistance to conscious control. These characteristics make them valuable indicators in emotional analysis research. While current micro-expression recognition technologies have made significant progress, their accuracy largely depends on correctly identifying the apex frame—the moment when emotional expression is most pronounced. Existing apex frame detection methods mainly rely on pre-annotated data or specific reference frames Though they perform well in controlled environments, these methods struggle with real-world challenges including multi-angle recording, lighting variations, and real-time processing requirements.
This paper proposes an innovative automated apex frame detection framework to address the limitations of existing technologies. We employ the high-performance RetinaFace model for facial feature extraction, coupled with an optimized 6×6 block segmentation method. This combination ensures stable facial feature localization under various shooting angles and lighting conditions. For apex frame detection, we implement a two-stage detection strategy. The first stage applies an improved uniform patterns LBP feature extraction technique that reduces dimensionality while enhancing recognition efficiency. This technique analyzes feature differences between all possible frame pairs rather than relying on a single reference frame, thereby identifying the pair with the most significant difference. The second stage calculates feature differences between each frame of the identified pair and all other frames in the video. It then employs a mode-based statistical principle to determine the true apex frame.
The main contribution of this research lies in avoiding the shortcomings of traditional methods that overly depend on the first frame or mean values. Our approach accurately captures the most intense expressions during micro-expression dynamics. Experimental results on SAMM, SMIC-E-long, and DAiSee datasets demonstrate that our method achieves an average 10% improvement in micro-expression recognition accuracy, and a 30% increase in processing speed compared to existing techniques. When processing high frame rate videos, our approach effectively reduces computational redundancy while maintaining high recognition precision. This method provides more reliable emotional recognition technology support for remote learning, online interviews, and telemedicine scenarios.
|
| 關鍵字： | |
| 編號： | 3 |
| 期刊或研討會名稱： |
第廿三屆離島資訊技術與應用研討會
台灣
|
| 著作名稱： | 基於多任務學習的信用卡詐欺偵測模型
|
| 年度： | 2025 |
| 類別： |
會議論文
|
| 摘要： | |
| 關鍵字： | |
| 編號： | 4 |
| 期刊或研討會名稱： |
第廿三屆離島資訊技術與應用研討會
台灣
|
| 著作名稱： | 利用強化學習優化計程車巡航路線推薦系統
|
| 年度： | 2025 |
| 類別： |
會議論文
|
| 摘要： | |
| 關鍵字： | |
| 編號： | 1 |
| 期刊或研討會名稱： |
IEEE Access
|
| 著作名稱： | Prompt Tuning Techniques for Chinese Idiom Recommendation
|
| 年度： | 2025 |
| 類別： |
期刊論文
|
| 摘要： | In this study, we propose a novel design incorporating six Planar Inverted-L Antennas (PILAs) to miniaturize Ultra-High Frequency Radio Frequency Identification (UHF-RFID) tags, thereby enhancing tag efficiency when mounted on various conductive surfaces. The design enables effective tag antenna operation when attached to surfaces with varying signal absorption and reflection characteristics, obviating the need for a complete antenna redesign. By dominating the capacitive reactance between gaps or the inductive reactance of vias, the tag matches the proposed antenna and the microchip (with an impedance of 10.7–j 134.5 Ω at 915 MHz). Both simulation and measurement results confirm the effectiveness of our approach. With this configuration of six adjacent coupled PILA antennas, the proposed tag minimizes electromagnetic field absorption from surfaces, such as containers filled with water or the human body, while enhancing radiation from reflective surfaces like metal. The tag antenna was fabricated on a single FR4 substrate, measuring 30×20×3.2 mm3 ( 0.10 λ0×0.08 λ0×0.01 λ0 ), and successfully tested on various conductive materials. In a free-space medium, a reader antenna with an Effective Isotropic Radiated Power (EIRP) of 4.0 W achieved maximum reading distances of 7.8 m, 6.1 m, and 5.3 m as the tag antenna was positioned on a metal plane, a water-filled container, and on a human body, respectively. |
| 關鍵字： | |