Face Detection and Localisation By Hesham Ahmed
Introduction � Aims � Different Methods used in Current Research � The Algorithm � Results � Demonstration � Limitations and Areas for Extension
Aims � To automatically detect if a colour image contains a human face and to locate that face within the image. � Cope with reasonable variations in terms of race, multiple faces, size, pose, illumination and background .
Different Approaches �������������� ����������� ������������� ��������������� ����� ������������������� ������������������ �������� ������� ���������!�������� Edges Feature Searching Colour
The Algorithm � Skin detection through colour. � Detection of eye/mouth regions through colour, edge detection and heuristics. � Ellipse estimation using Hough Transform � Weighting of Ellipse and Thresholding.
Skin Detection Using a YCbCr colour � transformation. Skin coloured pixels � form a tight cluster in Cb-Cr space. Involves a lot of � experimentation to find correct thresholds.
Edge-detection Firstly a Gaussian smoothing � mask is applied to remove spurious edges. Sobel operator used � (horizontal, vertical and both diagonals) Concentration of horizontal � edges used to detect eyes and mouths Other edges used with Hough � transform to calculate ellipses .
Feature detection Using combination of � Cb-Cr values and edge detection locate possible eye and mouth regions. Threshold values and � fix in blocks of 8x8 pixels.
Hough Transform and Ellipse Selection Hough Transform converts � points in Cartesian space into parametric space. For every point that would � fit a particular ellipse, it’s ‘accumulator’ cell would be incremented. Ellipses have 5 parameters � – extremely complex ( require 5 dimensional parameter space).
Results Better on images suitable for biometric � applications especially Face recognition. Images taken University College Dublin Colour � Face Database and personal collection of digital photos. Single face images, multiple face images, � different lighting, poses, accessories, races, orientation and occlusions.
Single face images ���������!������� #�������������������������� "����������������� "��������������� �����$��������������� %��������������������
More Faces ������!��� ���������!��� (������������������ &������$���������$�$��������� �������'
Multiple faces )� ������!��� �������������� *������������������ )� �����������������
Limitations Not particularly successful at detecting faces � in large groups (small faces). Due to effects of noise , thresholds , unclear � features and bias of Hough Transform .
Areas for future extension � Handle rotated faces by looking for eye-candidates in several directions. � Develop greater sophistication in confirming/rejecting face candidates � Locate features with greater accuracy.
Problems encountered and solutions Lack of detail of implementations � Derived new and own ways of achieving the task e.g. � experimenting with thresholds, formulae and techniques. Complexity of Hough Transform � Restricted to 2D –parameter space by estimating origin � from eye candidates and only considering of upright ellipses of specified proportions. Also counted one in every two possible edges to improve � speed.
Demonstration…
Conclusions Illustrated that one can still achieve good results � with a fraction of the computation and a simple algorithm. Identified weaknesses in other existing research. � Also merged ideas and experimented with different � ways of achieving face detection( including own equations for colour feature detection).
Recommend
More recommend