全國中小學科展

2025年

自監督學習在臺灣手語辨識上之應用研究

在臺灣手語辨識,先前研究所使用的監督式學習需要大量標記樣本而限制可辨識詞彙量。為此,本研究借鑒自然語言處理領域中BERT 的遮罩想法,將未標記手語影片隨機遮蓋部分幀數,並讓模型學習預測被遮蓋的幀數以學習臺灣手語的特徵,並透過遷移學習來訓練辨識模型,此作法可克服現有臺灣手語資料缺少的問題。經過實驗,本研究訓練之詞彙辨識模型達成了242 個詞彙量,94.8%的準確率。 此外,先前研究皆未在手語句子翻譯上有成果。因此本研究基於預訓練模型,整合設計手語翻譯的系統,實驗中,系統在100 個句子的翻譯表現達到88%的準,且BLEU-4 分數取得20.98,證明自監督學習的方式在手語辨識、翻譯上是有效的。並展現出樣本需求少與辨識詞彙量可輕易擴大的潛力。

數位物理實驗室:毫米波雷達系統之設計與應用

本研究旨在設計基於毫米波雷達的數位物理實驗系統,用於精確量化彈簧簡諧運動。傳統物理實驗易受肉眼觀察與手動測量的誤差影響,本系統利用24GHz毫米波雷達結合自製電路板,進行即時、無接觸的運動測量。透過設計電路板、撰寫韌體訊號轉換程式,並進行數位數據分析,成功開發了靈敏的毫米波雷達系統。我們利用彈簧簡諧運動實驗驗證了該系統,觀察不同質量砝碼對彈簧運動頻率的影響。實驗結果顯示,考慮彈簧質量後,測量數據與理論結果的均方根誤差從0.62Hz降低至0.35Hz,顯示出系統的高度精確性及穩定性。本研究成功解決了傳統實驗中的量測誤差問題,以毫米波雷達技術實現了精確觀測。開源設計有助於推廣至學校的物理實驗室,為學生提供先進的實驗工具與數據分析經驗。這展示了毫米波雷達在物理實驗中的應用潛力,並為未來教學實驗提供了高效、低成本的解決方案。

Greenhouse Gases Reduction: Conversion of Methane and Carbon Dioxide into Clean Energy

In the upcoming years, both population and energy consumption are expected to increase dramatically [1]. Industrialization has led to a dramatic shift in the energy environment [2], with predictions of a 57% increase in demand for energy between 2002 and 2025 [3]. In addition to organic materials like trees and solid waste, fossil fuels like coal, natural gas, and oil provide more than 90% of the world's energy needs. Their overuse has resulted in the release of climate-altering greenhouse gases like carbon dioxide (CO2) and methane (CH4) into the atmosphere [4]. Scientists and other stakeholders are putting more emphasis on finding solutions to global warming, increasing energy production in order to meet increasing demands, and decreasing emissions of greenhouse gases. Using greenhouse gasses to make useful chemicals or fuels is one solution to both problems [5]. This motivated researchers to investigate the potential of CO2 and CH4 as clean energy sources. The process of dry reforming of methane (DRM) has been identified as a potentially successful strategy for transforming CO2 into marketable syngas with a balanced H2/CO composition [6], [7], [8], [9]. The economic viability of DRM, the reactor type, the availability of raw materials, and the intended use of the produced syngas are all-important considerations. Though DRM is gaining popularity, maintaining its long-term stability is difficult due to carbon accumulation from CO disproportionation and methane degradation [10], [11]. The catalyst used, as well as other parameters like as pressure, temperature, feed concentration, and reactor size, are critical to the process's effectiveness. In this scenario, a nickel catalyst on a La2O3/SiO2 substrate with microspheres and a core-shell structure will be developed to improve the conversion of greenhouse gases into profitable syngas. This catalyst is projected to improve the efficiency and performance of the DRM process significantly.

自監督學習在臺灣手語辨識上之應用研究

在臺灣手語辨識,先前研究所使用的監督式學習需要大量標記樣本而限制可辨識詞彙量。為此,本研究借鑒自然語言處理領域中BERT 的遮罩想法,將未標記手語影片隨機遮蓋部分幀數,並讓模型學習預測被遮蓋的幀數以學習臺灣手語的特徵,並透過遷移學習來訓練辨識模型,此作法可克服現有臺灣手語資料缺少的問題。經過實驗,本研究訓練之詞彙辨識模型達成了242 個詞彙量,94.8%的準確率。 此外,先前研究皆未在手語句子翻譯上有成果。因此本研究基於預訓練模型,整合設計手語翻譯的系統,實驗中,系統在100 個句子的翻譯表現達到88%的準,且BLEU-4 分數取得20.98,證明自監督學習的方式在手語辨識、翻譯上是有效的。並展現出樣本需求少與辨識詞彙量可輕易擴大的潛力。

矩形密鋪及其應用

「在格狀平面中用矩形以互不重疊的方式鋪滿(2D rectangle tiling problem)」為一NP-complete問題(Dani`ele Beauquier et al ,1995),目前多項式時間只能求出盡可能覆蓋最大面積的近似解。本研究所創的階梯演算法 stair algorithm 透過改變動態規劃紀錄狀態的方式,使狀態數大幅減少,進而改善求準確解的時間複雜度,也成功證明此演算法的正確性。本研究的演算法可被應用於平行計算中的負載平衡、積體電路設計等方面。隨後,本研究寫了一個互動展示品清楚呈現此演算法的功能。且以階梯演算法成功檢驗並比較 RTILE PROBLEM 的 7/3-approximation algorithm (Krzysztof Lorys and Katarzyna E. Paluch,2000 [4]) 與 11/5-approximation algorithm (Piotr Berman et al,2001[7])進行比較與分析。

探討濕地耐鹽菌對植物耐鹽及根部的交互作用

本研究從濕地篩選出可能為新種的耐鹽菌Oceanobacillus sp.,暫命名為OC2,其在無植物相伴狀態下不會降低土壤含鹽量,但卻在與植物共存後誘發特殊機轉,促使土壤含鹽量降低約,並提升植物耐鹽能力,顯示 OC2與植物存在特殊交互作用。深入研究發現,OC2能產生IAA,並吸引植物根部向其生長以利其進入根部,並在鹽逆境下分泌代謝物以刺激植物合成脯胺酸 (增加達98.5%)提升根部滲透壓、增加葉片類胡蘿蔔素及類黃酮含量以提升植物抗氧化力。植物方面,鹽逆境下植物分泌的化學物質會觸發OC2產生更多的IAA(約17%),藉以刺激植物根系發展以利水分吸收,而OC2的存在會促進根部澱粉酶活性上升達88%,以分解澱粉產生可溶性醣類供OC2使用,推測兩者存在共生關係。本研究展示新種耐鹽菌與植物的交互作用,期待透過此菌改善鹽化農地並能提升作物產量。

Real-Time Ensemble Model for Stroke, Drowsy, and Distracted Driver Detection Using Transfer Learning Models

Road safety remains a global concern, with driver-related factors like distraction, drowsiness, and medical conditions such as stroke being leading causes of accidents. In this paper, we propose a real-time ensemble learning framework that leverages transfer learning for the detection of stroke, drowsiness, and distracted driving. Our model integrates multiple Convolutional Neural Networks (CNNs) fine-tuned for each specific task, and employs a stacking method to combine the predictions of these models using a meta-classifier. Notably, the model is optimized to enhance stroke detection, minimizing false negatives— an essential aspect for timely medical intervention. Experimental evaluations on diverse datasets demonstrate the efficacy of our approach, achieving an overall accuracy of 92.5%. The results emphasize the model’s potential for real-time driver monitoring, offering critical safety features that could reduce accidents and save lives.

Application of Carbon Aerogels in Lithium-Air Batteries

One of the main challenges with today’s batteries is their relatively low volumetric and specific capacities. The highest specific capacity can be achieved with lithium-air batteries, which use metallic lithium as the anode and typically some form of porous carbon as the cathode. To enhance performance, aerogels—among the world’s lightest solid materials—are ideal candidates for cathodes. Resorcinol-formaldehyde (RF)-based carbon aerogels, for example, serve this purpose well. In my work, I utilized two types of carbon aerogels as cathode materials: one derived from pyrolyzed resorcinol-formaldehyde polymer and the other a graphene-oxide-modified version of this carbon gel. I integrated the carbon aerogels I had pyrolyzed into lithium-air batteries to improve the cell’s performance, energy density, and capacity compared to cells using activated carbon. In my research, I examined the pore structure and surface properties of these materials in aqueous media using NMR (nuclear magnetic resonance) relaxometry and cryoporometry, exploring their impact on battery efficiency. I found that the graphene-oxide-containing sample's pores filled with water in a layered manner, indicating a more hydrophilic surface, which suggests a denser arrangement of oxygen-containing functional groups compared to the unmodified carbon aerogel. The pore sizes were reduced after adding graphene oxide, resulting in an increased specific surface area for the sample. Incorporating the reduced graphene-oxide-containing carbon aerogel enabled the creation of a more efficient, higher-capacity battery than with the RF carbon aerogel. This improved performance is likely due to the aerogel’s higher oxygen content and altered morphology. The increased oxygen content provides more active sites for oxygen reduction, meaning that a greater specific power output can be obtained from the battery.

短期睡眠剝奪對小鼠免疫系統的影響

現代社會中,睡眠剝奪已成為普遍問題,人們對其對免疫系統及整體健康的負面影響愈加關注。本研究使用特製的旋轉鼠籠讓小鼠連續72小時保持清醒,探討急性睡眠剝奪對小鼠免疫反應的影響。研究發現NK細胞與脾臟中的記憶CD8 T細胞比例明顯減少,顯示細胞毒性功能受損或記憶免疫反應下降。與此同時,抗炎細胞因子的表達增加,而促炎細胞因子和相關基因的表達則有顯著下調。此外,雖然觀察到B細胞比例有所增加,這可能是免疫系統在細胞免疫功能受損時,維持免疫穩態的反應。這些發現揭示了睡眠剝奪可能抑制免疫系統造成損害。本研究強調適量睡眠對維持免疫平衡的重要性,並指出睡眠不足可能促進慢性免疫問題的發展。在此基礎上,後續研究可探討短期睡眠剝奪與腫瘤及免疫系統的關聯,並延伸至長期剝奪的影響。

二氧化鋯量子點在文物修復與減碳科技應用的潛力

本研究成功以水熱法在 110°C 下合成了約3.90 nm 大小的ZrO2量子點(QDs)。此設計的ZrO2 QDs 能隙為5.03 eV(波長λ < 300 nm),在可見光和紫外光範圍內無明顯吸收特徵,呈現高度惰性和穩定性,適合應用於抗紫外線塗層或顏料。而ZrO2 QDs 表面豐富的氧空位與不同溫度下的CO₂轉化率及CO/CH₄產物選擇性相關。氧空位為帶部分正電的酸性活性位,CO2作電子受體為路易士酸。經氧氣環境加熱處理後的ZrO2 QDs 能提高CO2轉化率且在低溫條件下選擇性較高能促進電子轉移生成CH₄(每分子8e⁻ 轉移)。不同金屬簇(如Fe、Ni、Co和Cu)表面修飾後,Fe-ZrO2 QDs 被證明為最佳催化劑,低溫下更有效促進CH₄生成,且優於ZrO2 QDs。這顯示Fe與ZrO2間存在顯著的強金屬-載體相互作用(SMSI),提升Fe捕捉CO₂分子的能力。此特性突顯ZrO2於碳減排技術的潛力,能有效將CO₂轉化為可再利用的碳基燃料或化學原料,為減少溫室氣體提供實用解決方案。