- 學位論文
運用人工智慧偵測方法增強機場安全：結合不同X光能量之先導研究
Enhancing Airport Security through AI-based Detection：A Pilot Study Combining Different X-ray Energy
摘要
目的：行李檢查對航空安全至關重要，而傳統的檢查方法需要行李檢查員檢視大量的X光圖像，導致準確性和效率低下，本研究旨在通過開發一個自動及高準確度的航空行李檢查識別系統來應對這些挑戰，該系統融合了多能量X光掃描和人工智能技術 (AI)，提高了乘客安全性和查緝走私的效果。 材料與方法：在這項先導研究中，我們造影了26張常見走私物品的X光圖像，如大蒜、香菇、豬肉製品、菸 (紙菸及加熱菸) 和酒，這些圖像使用60、80及120 kV的管電壓、10 mA的管電流和100/600毫秒的曝光時間進行造影，然後，我們將三種不同能量的圖像合成為一張單一的X光圖像，並使用RGB顏色模型對這些圖像進行著色，這些圖像被標記為多能量圖像組 (JPGH)，另一組圖像，直接將其偽色彩化，形成了偽色彩圖像組 (JPGC)，為了訓練YOLOv4模型，這些圖像通過水平和垂直翻轉後，再調整影像強度進行了資料擴增，總共生成了1,170張擴增圖像，最後，模型被訓練來自動識別JPGH和JPGC組中的有機 (O) 和非有機 (NO) 物品，並使用命中率 (Hit Rate) 和交疊率 (Intersection Over Union , IoU) 兩種指標來驗證效能。 結果：JPGH組中偵測有機物和非有機物的命中數和命中率分別比JPGC組好2.038和1.200倍，JPGH組中有機物和非有機物的交疊率分別比JPGC組好1.090和1.046倍。因此，多能量X光圖像在經過完全卷積網絡 (FCN) AI模型處理時，顯示出更優越的識別性能和準確性。 結論：將AI技術與多能量X光圖像融合，為行李檢查提供了更有效和準確的方法。經過良好訓練的深度學習演算法能夠準確識別危險和違禁物品，從而降低失誤的比率，這項先導性的研究為未來將AI應用在增強航空安全方面奠定了基礎。
並列摘要
Purpose： Baggage inspection is critical for ensuring aviation safety. Traditional inspection methods require human inspectors to screen a large volume of X-ray images, leading to low accuracy and efficiency. This study aims to address these challenges by developing an automated, high-accuracy recognition system for aviation baggage inspection. This system integrates multi-energy X-ray scanning with artificial intelligence (AI) techniques, enhancing both passenger safety and anti-smuggling efforts. Materials and Methods： In this pilot study, we acquired twenty-six X-ray images of items commonly smuggled, such as garlic, mushrooms, pork products, tobacco, and alcohol. The images were captured using X-ray parameters of 60, 80, 120 kV, a tube current of 10 mA, and exposure times of 100 and 600 msec. We synthesized three different energy images into a single X-ray image, which was colorized using the RGB color model; these images were labeled as the multi-energy image group (JPGH). Another set of images, directly pseudo-colored, formed the pseudo-color image group (JPGC). To train the YOLOv4 model, these images were augmented through horizontal and vertical flipping and intensity modulation, resulting in 1170 augmented images. The model was then trained to automatically recognize organic (O) and non-organic (NO) goods in both the JPGH and JPGC groups. Metrics such as hit rate and intersection over union (IoU) were employed for validation. Results： The hit and Hit Rate for organic and non-organic goods in the JPGH group were 2.038 and 1.200 times better, respectively, compared to the JPGC group. Additionally, the IoU for organic and non-organic goods in the JPGH group was 1.090 and 1.046 times better, respectively, than that in the JPGC group. Hence, multi-energy X-ray images showed superior recognition performance and accuracy when processed through Fully Convolutional Network (FCN) AI models. Conclusion： The integration of AI techniques with multi-energy X-ray images offers a more efficient and accurate approach to aviation security inspection. Well-trained deep learning algorithms can accurately identify dangerous and prohibited items, thereby reducing the rates of false positives and negatives. This pioneering research serves as a foundation for future applications of AI in enhancing aviation security.