University of Twente Student Theses


LULU operators for image segmentation and object detection

Nigatu, Amare Degefaw (2013) LULU operators for image segmentation and object detection.

[img] PDF
Abstract:Image segmentation remains a challenge task; most of the segmentation approaches may not be even applicable in remotely sensed image because of the complexity due to multi-spectral, multi-scale and heterogeneity properties. The aim of this study is to segment and automatically identify objects based on LULU operators recursive application, the Discrete Pulse Transform and the scale space analysis. The LULU-DPT algorithm, The Discrete Pulse Transform 2D algorithm and the scale-space automatic object identification algorithm are applied on a very high resolution image on urban areas of Cairo city, Egypt. Six areas are selected from different parts of the city. The first principal component analysis is applied and the subset images are decomposed by recursive application into one dimension and two dimensions discreet pulses. From man-made features, object of interest such as buildings and roads are automatically identified from pulses based on the scale space image analysis that generates the breaking points at which objects are drastically changed in size and shape, change in mean and standard deviation along the scale function. In multi resolution analysis, segments and objects are identified from the decomposed input image using DPT either partial reconstruction from selected pulses by identifying feature of interest and removing back ground features or fully reconstruction from the sum of all non-zero DPT levels. Identified objects are compared with outputs of thresholding segmentation method, complex urban features such as buildings and apartment blocks are easily identified by pulses. Segmentation accuracy assessment is resulted a high accuracy output; with area fit index for buildings 0.10 to 0.11 and for road 0.19 to 0.21. The maximum overlap between the reference data and segmented output is 90% for building and 81% for roads. The object boundary fit between the segmented object and reference is 0.7 after the correction factor 0.90 obtained and applied. Keywords: LULU operators, DPT, MRA, scale-space, Area Fit Index
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page