基於頻譜影像之空氣汙染檢測方法

專利類型

發明

專利國別 (專利申請國家)

中華民國

專利申請案號

110109760

專利證號

I 777458

專利獲證名稱

基於頻譜影像之空氣汙染檢測方法

專利所屬機關 (申請機關)

國立中正大學

獲證日期

2022/09/11

技術說明

近年來空氣汙染已經成為全世界面臨解決的問題，尤其粒徑小於2.5 微米的顆粒物(PM2.5)能透過空氣傳播將有害的物質藉由人類的呼吸到肺部，因而產生嚴重的健康問題。 本研究結合超頻譜影像技術和深度學習，提出了一種大範圍、低成本的空氣汙染檢測之方法，透過對空拍機進行可見光之超頻譜影像技術之建模，將空拍機拍攝之影像賦予超頻譜資訊；在分類演算法部分，分別提出了三維卷積神經網路自編碼(3D Convolutional Neural Network Auto Encoder)以及主成分分析法（Principal components analysis，PCA）並結合 VGG-16 (Visual Geometry Group)以探討空 氣汙染的光學特性與模型識別的準確性。 結果顯示在混淆矩陣中 RGB 圖像完全分類正確的有 210 張pm2.5_good、168 張 pm2.5_moderate 和 149 張 pm2.5_severe，PCA 圖像完全分類正確的有 242 張 pm2.5_good、115 張 pm2.5_moderate 和178 張 pm2.5_severe，3D-CAE 圖像中完全分類正確的有 15 張pm2.5_good、80 張 pm2.5_moderate 和 103 張 pm2.5_severe，在平均分類準確率中 PCA+VGG-16 有最好的分類精度，RGB+VGG-16 次之，最後是 3D-CAE+VGG-16。 In recent years, air pollution has become a problem facing the world. Particularly, particulate matter (PM2.5) with a diameter of less than 2.5 microns can spread through the air and bring harmful substances to the lungs through human breathing, thus causing serious health problems. This study combines hyperspectral imaging technology and deep learning to propose a large-scale, low-cost air pollution detection method. Through the modeling of the visible light hyperspectral imaging technology of the aerial camera , the image captured by the drone camera is given Hyperspectral information . In the classification algorithm part, 3D Convolutional Neural Network Auto Encoder and Principal Components Analysis (PCA) are respectively proposed and combined with VGG-16 (Visual Geometry Group) to Discuss the optical characteristics of air pollution and the accuracy of model recognition. The results show that 210 of pm2.5_good, 168 pm2.5_moderate, and 149 pm2.5_severe images are completely classified as RGB images in the confusion matrix, and 242 pm2.5_good and 115 pm2 images are fully classified as PCA images. 5_moderate and 178 pm2.5_severe. Among the 3D-CAE images, 15 pm2.5_good, 80 pm2.5_moderate and 103 pm2.5_severe are completely classified correctly. Among the average classification accuracy rates, PCA +VGG-16 has the best Good classification accuracy is followed by RGB+VGG-16, and finally 3D-CAE+VGG-16.

備註

連絡單位 (專責單位/部門名稱)

技術移轉授權中心

連絡電話

05-2720411轉16001

網址