以影像特徵自動化估計影像模糊指標

Abstract

無人飛行載具(Unmanned Aerial Vehicle, UAV)為一種新興的遙測技術，其具有低成本、高機動性等優勢，可在短時間內拍攝大量高解析度、高重疊率影像，被廣泛應用於諸多領域。以無人飛行載具進行小範圍高精度製圖對於影像品質要求甚高，影像品質將直接影響產品精度，而影像模糊是小型無人飛行載具影像常見的問題之一。影像模糊主要成因為，拍照時受到側風或亂流影響，傳統上，由於影像數量很多，大多忽略其影響，或者以人工挑選出模糊影像，但人工挑選耗時耗力，且模糊程度仰賴人為判斷，沒有固定的標準，若能訂定影像模糊指標，即可客觀評估各影像之品質，從中挑選清晰影像，並剔除模糊影像，作為後製產品之輸入影像，在一定程度上能提升產品精度。 本研究提出以影像特徵自動化估計影像模糊指標，所採用之影像特徵基於點散函數及頻率域分析，分別擬訂兩個評估都市區及森林區影像之模糊指標，最後配合非監督式分類，達成自動化判識模糊影像，由分類結果分析比較此二指標之效益。研究主要步驟包含：特徵萃取、模糊指標擬定及影像分類。在特徵萃取方面，點散函數指標萃取直線特徵，並計算其點散函數；頻率域指標則萃取影像頻譜圖之高頻資訊。模糊指標擬定方面，點散函數指標採用各方向點散函數分布擬合橢圓，其面積及均向性分別代表模糊強度指標及模糊方向性指標；頻率域則採高頻資訊幾何特徵擬合橢圓，但僅以均向性作為模糊方向性指標。在影像分類方面，採用K-means非監督式分類，點散函數特徵參考兩個特徵；頻率域特徵由於僅參考單一特徵，故使用Otsu計算門檻並分類。 本研究使用資料分為城市區UAV影像、森林區UAV影像及運動型相機錄像，實驗區域分別為國立交通大學光復校區、鹿場及國立交通大學工程二館一樓。在影像分類方面，本研究提出之點散函數特徵模糊指標，在交大光復校區及工程二館之分類精度分別可達71%及90%；頻率域特徵模糊指標則用於鹿場影像，挑選清晰及模糊影像序列，研究成果顯示，在正射影像的解算中，若由模糊影像序列改為清晰影像序列，其成功較列率約提升6%，達到70%，顯示使用清晰影像對於攝影測量帶來益處。最後本研究擬定了點散函數之多尺度策略，初步驗證該策略將使合格線篩選成功率提升73%，使點散函數特徵指標更為完善。

Unmanned Aerial Vehicle (UAV) is an advance Remote Sensing technology in data acquisition. UAV has many advantages such as low cost, high flexibility with lots of high spatial resolution and high overlapped images. UAV has been widely used in many fields, in which the image quality plays an important role in UAV’s applications. However, the small wings-fixed UAV is suffering from the image blur due to crosswind and turbulence, degrading the production quality. Traditionally, most of them ignore the influences of blurred images due to a large number of images, and it is labor-intensive to identify the blurred images manually. But it is not suitable for manual identification without a stationary standard. If a reliable blur index for the image is available, it’s possible to estimate the quality of each image automatically. Only selected sharp images from all images will be included in photogrammetry process. In this study, we proposed an automatic blur index extraction using image-derived features. The image-derived features are Point Spread Function (PSF) and frequency domain which are used to develop blur indexes for urban and forest areas independently. K-means classifier was adopted for image classification by the blur index without any manual identification and analyzes these two classification results. The major steps include: feature extraction, blur index formulation and image classification. In feature extraction, PSF-derived blur index extracts linear features and estimate the PSF; Frequency-derived blur index extracts the high frequency of the image spectrum. In blur index formulation, PSF-derived blur index fits ellipse by the distribution of all PSFs. The area and isotropy of the ellipse represent the intensity and direction of blur, respectively; Frequency-derived blur index fits ellipse by the geometric features of high frequency. The isotropy of the ellipse represents the direction of blur. In image classification, we use unsupervised classification. PSF-derived blur index uses k-means clustering with two features; Frequency-derived blur index uses Otsu to calculate the threshold for image classification. The experimental data includes urban UAV images, forest UAV images and 4K video from GoPro action camera. They are located on National Chiao Tung University (NCTU) Kuangfu campus, Deer field and NCTU Engineering Building 2 respectively. In image classification, we used the proposed PSF-derived blur index to classify the blurred and sharp images, and obtained the overall accuracies of 71% and 90% in urban UAV images and 4K video, respectively. Frequency-derived blur index is used in forest UAV images for the selection of sharp and the blurred images sequence. The experimental results revealed that the sharp image sequence is better than blurred image sequence in the generation of orthophoto image. The improvement rate for the alignment rate is 6 %, and the sharp image sequence alignment rate reachs 70 %. The experimental results revealed that sharp images could benefit photogrammetry in product quality. Finally, this study has developed a multi-scale strategy for the extraction of PSF. The preliminary verification of this strategy will improve the success rate of qualified lines by 73%, so that makes PSF-derived blur index be more robust.