Kernel Regression Based Technique For False Edge Elimination
Loading...
Date
2014-07
Authors
Abu Samah, Hafizi
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
Edge detection is the front end processing step in many computer vision applications. It
is used to detect edges corresponding to the boundaries of objects contained in an image.
Edge detection simplifies the image by filtering out the unwanted information and remained
the significant information. The main challenge in edge detection is to reduce
the effect of false edge detection. A false edge is usually detected on texture or noise.
The current edge detection techniques are bound to certain limitations which lead to unsatisfactory
result of edge detection. Furthermore, the current edge detection techniques
can be considered as pre-processing techniques because the efforts of reducing the false
edge detection are conducted prior to the edge detection. The pre-processing techniques
are sensitive toward the changes of the original image or image gradient which can affect
the locality of the detected edge. Thus, a post-processing technique of edge detection
is proposed in this study. The proposed method is a technique of discriminating the
true and false edges in an already detected edge image. The proposed method employs
locally adaptive regression kernel (LARK) as feature descriptors and applies the concept
of the LARK object and face detection algorithm to detect false edges. By using LARK,
false edges are identified by comparing the detected edges’ LARK with the reference
flat region kernel descriptor. False edges are detected if the edges’ LARK is similar to
the reference flat region kernel descriptor. The detected false edges will be eliminated
while the remaining edges are considered as the true edges. Three types of reference flat
region kernel descriptor: Small, Medium and Large were proposed in this thesis. Each
type of the reference flat region kernel descriptor is capable to eliminate false edges on
different types of textured region. False edge elimination is optimized by combining the
proposed reference flat region kernel descriptors. The reference flat region kernel descriptors
were combined into three combinations: Small and Medium; Small and Large;
Medium and Large. The proposed method was tested on images obtained from Berkeley
segmentation dataset (BSDS) and University of Gronigen (RuG) dataset. The proposed
method was evaluated qualitatively and quantitatively and the results were compared
with other edge detection techniques. The qualitative evaluation result shows that the
proposed method with Small and Medium combination produced edge images which
contain few false edges and the images are better in comparison to other edge detection
techniques. For qualitative evaluation, the proposed method has also produced the best
results with average Baddeley’s delta metric and average F-measure, 23.37 and 0.2209,
respectively.
Description
Keywords
Kernel regression based technique , for false edge elimination