無人載具遙測影像之移動模糊偵測及影像修復

Detection and restoration of motion blur for UAV imagery

詹凱智(Kai-Zhi Zhan)

指導教授 ： 張智安

國立交通大學/工學院/土木工程系所/碩士(2015年)

https://doi.org/10.6842/NCTU.2015.00834

摘要

無人飛行載具(Unmanned Aerial Vehicle, UAV)為一種新興的遙測技術(Remote Sensing, RS)，無人飛行載具具有低成本、高機動性及拍攝大量高解析度、高重疊率之影像等優勢，近年來被大量應用於許多領域，如：三維都市建模、緊急救援、變遷偵測等等。其中，小範圍高精度製圖對於影像品質要求甚高，影像模糊是小型定翼無人飛行載具常見的問題之一，影像品質與產品精度有直接的關聯。造成影像模糊的原因主要為拍照時有側風或亂流影響。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，但人工挑選太過耗時耗力，且模糊程度需仰賴人為判斷，若能以自動方式偵測影像模糊，並進行修復，在一定程度上能提升產品精度。
本研究提出以無人飛行載具上搭載之定向及定位系統(Position and Orientation System, POS)輔助，偵測無人飛行載具影像中之模糊影像，並探討模糊程度對於攝影測量造成之影響。研究步驟包含：模糊影像偵測、模糊影像修復及驗證分析。在模糊偵測方面，研究中改進傳統的線性運動模糊指標(blur in pixels)，提出以共線條件式為主體，計算無人飛行載具運動時之模糊程度(degree of motion blur)。偵測模糊影像方法使用支持向量機分類器(Support Vector Machines, SVM)，分類特徵除無人飛行載具之模糊程度外，加入影像資訊、定向定位系統資料及線性運動模糊指標。除無人飛行載具影像外，研究中採用小型線性振動台模擬無人飛行載具線性運動行為，並討論影響模糊程度之可能影響因子。模糊影像修復方面，本研究採用露西-理查德森反卷積法(Lucy –Richardson deconvolution)修復模糊影像，而修復影像需先定義影像之點散函數(Point Spread Function, PSF)或稱模糊核(blur kernel)，研究中使用運動模糊程度作為模糊核大小與角度給定之依據。在驗證分析方面，主要分為兩部分討論：特徵萃取與影像匹配。特徵萃取以加速穩健特徵(Speeded up robust features, SURF)萃取點特徵，Canny 邊緣線偵測法(Canny edge detector)萃取線特徵。影像匹配方面，採用加速穩健特徵匹配。
本研究使用資料為SenseFly eBee UAV系統，實驗區域為國立交通大學光復校區，實驗影像為129張。模糊影像偵測方面，使用本研究提出之運動模糊指標，分類精度可達76%。研究成果顯示，影像模糊程度越大確實會減少特徵萃取的數量，以及影像匹配對數與成功率。在影像修復後，可以增加特徵萃取數量，點特徵增加量約57%，線特徵增加量約30%。於影像匹配方面，雖然影像修復後，匹配對數減少，但匹配正確率上升。顯示影像修復對於攝影測量帶來益處。

關鍵字

無人飛行載具 ； 模糊影像偵測 ； 模糊影像修復 ； 運動模糊程度

並列摘要

Unmanned Aerial Vehicle (UAV) is an advance Remote Sensing (RS) technology in data acquisition. UAV has many advantages such as low cost, high flexibility, high spatial resolution and high overlapped images. UAV has been widely used in many fields such as mapping, 3D city modeling, emergency rescue, change detection and etc., in which the image quality plays an important role of 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, small-frame UAV requires a large number of images for processing, and it is labor-intensive to identify the blurred images manually. The automation of blurred image detection and restoration will save processing time and improve final product qualities of UAV.
In this study, we proposed a Position and Orientation System (POS) assisted method to detect the blurred image, and discussed the influence of motion blur images for photogrammetric purposes. The major steps include: blurred image detection, restoration and verification. In blurred image detection, we modified the traditional degree-of-linear-blur (blinear) method to degree-of-motion-blur (bmotion) based on the collinear condition equation to describe the degree of blur of UAV images. Support Vector Machines (SVM) classifier was adopted for blur detection and feature extraction (e.g., image information, POS data, blinear and bmotion). In addition, we used a shaking table to simulate the linear motion, and discussed the impact factors of image blurs. Lucy–Richardson deconvolution was to conduct blurred image restoration. The kernel size and the rotation angle of Point Spread Function (PSF) (also known as blur kernel) were defined by degree-of-motion-blur.
In verification, we explored the impacts of blurs from two aspects: feature extraction and image matching. In feature extraction, we applied Speeded up robust features (SURF) and Canny Edge Detector to extract the feature points and lines, respectively. In image matching, we used SURF matching to evaluate the influence of image blurs.
The experiment was performed using SenseFlyeBee UAV system with the location on Kuangfu campus, National Chiao Tung University (NCTU). The total number of image taken was 129. In blurred image detection, we used the proposed degree-of-motion-blur as a feature to classify the blurred and sharp images, and obtained the overall accuracy of 76%. Besides, the number of features and the image matching success rate were associated with the degree of blur. Higher degree of blur hindered number of extracted features and image matching rate. The number of features and the image matching success rate increased after blur restoration. The improvement rates for feature point and line were 57% and 30%, respectively. In image matching, although the number of matching pair reduced, the correctness increased. The experimental results revealed that blurred image restoration can benefit photogrammetry in product quality and processing time

並列關鍵字

Unmanned Aerial Vehicle ； blur image detection ； blurred image restoration ； degree of motion blur

參考文獻

Chen, C., Chen, W., and Bloom, J.A., 2011. A universal reference-free blurriness measure, IS&T/SPIE Electronic Imaging. International Society for Optics and Photonics, pp. 78670B-78670B-78614.

Duan, F., and Zhang, Y., 2009. The estimation of blur based on image information, IEEE International Conference on Image and Graphics (ICIG), pp.109-112.

Addiati, D., Sheikh, U.U., and Bakar, S.A., 2009. Restoration of out of focus barcode images using Wiener filter, IEEE Conference on Signal-Image Technology and Internet-Based Systems (SITIS), pp. 133-137.

Cai, J.F., Ji, H., Liu, C., and Shen, Z., 2009. Blind motion deblurring from a single image using sparse approximation, IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 104-111.

Canny, J., 1986. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 8(6): 679-698.