6493

6057

كارشناسي ارشد

مخابرات

اصفهان: دانشگاه صنعتي اصفهان، دانشكده برق و كامپيوتر

1390

نه، 89ص.: مصور، جدول، نمودار

ص.ع. به فارسي و انگليسي

نيلوفر قيصري، رسول امير فتاحي

سعيد صدري

2/12/90

بهزاد نظري، حسين رباني

1396/10/12

كتابنامه

برق و كامپيوتر

مهندسي برق و كامپيوتر

ID6057

به فارسي و انگليسي: قابل رويت در نسخه ديجيتالي

Semantic Segmentation of Aerial Images Using Features of Color and Texture Mohaddeseh Ghiasi m ghiasifathabad@ec iut ac ir Date of Submission 20 09 2011 Department of Electrical and Computer Engineering Isfahan University of Technology Isfahan 84156 83111 Iran Degree M Sc Language PersianSupervisors Dr Niloofar Ghaisari n gheissari @cc iut ac ir Dr Rasul Amirfattahi fattahi@cc iut ac irAbstractIn this thesis a semantic segmentation method for aerial images is presented Semantic segmentation allowsthe task of segmentation and classification to be performed simultaneously in a single efficient step In fact a semantic label is assigned to each segment which identifies its category such as tree road buildings grass and water This is an important aspect of an autonomous landing system However for anautonomous aircraft a mere segmentation is not sufficient We also require a semantic understanding of theinput image to assess the possibility of a safe landing Combing these two stages is the main motivation ofapplying a semantic segmentation method This algorithm relies on descriptors of colour and texture In thetraining phase we first manually extract homogenous areas and label each area semantically Then colourand texture descriptors for each area in the training images are computed The pool of descriptors and theirsemantic label are used to build a classifier KNN To segment a new image we over segment it into anumber of super pixels Super pixels provide an automatic uniform and homogenous splitting of the inputimage Then we compute texture or colour descriptors for each super pixel and choose the kth nearestneighbour to this multi dimensional vector from the KNN This labels the superpixel semantically Labelling all super pixels provides a segmentation map We used LBP HF and colour histograms of RGBimages as texture and colour descriptors respectively This decision is based on a thorough evaluation studyon different texture descriptors as stated in this thesis LBP HF is globally rotation invariant and has beenproved to outperform many state of the art texture descriptors This algorithm is applied to a large set ofareal images and is proved to have aboute 96 success rate As boundaries of super pixels coincide withthe image edges the algorithm is capable of preserving boundaries Occluded areas are successfullysegmented and labelled similar Since the segmentation is performed in super pixel level not the pixellevel the test phase is very fast and effective This is a very important advantage of our algorithm Thismethod has been also compared against similar methods on the same database The comparison resultsconfirm the superiority of the proposed method Keywords Semantic Segmentation Texture descriptor aerial images Oversegmentation superpixel

نيلوفر قيصري، رسول امير فتاحي

سعيد صدري

بهزاد نظري، حسين رباني