Discipline: Computer Sciences and Information Management
Subcategory: Computer Science & Information Systems
Yahkeef Davis - Onondaga Community College
Co-Author(s): Vicentica Valdes, Onondaga Community College, Syracuse, NY
The CSTEP/LSAMP Office at Onondaga Community College has continued to develop facial recognition software program using the MatLab Computer Vision System toolbox to automate collection of data used to both report on and better allocate program resources.
Core components of the software had been identified (image acquisition, training of the classifier, identification of the query image and recording data, to include the development of the database used to train the classifier) by previous groups and included an algorithm partially coded in MatLab. The algorithm was modified and coding in MatLab was completed. Testing of the software was then done with a small database, primarily working to ensure the algorithm was correctly coded. The database consisted of seven students with eight images of each student; testing was done using images from the database. Successful coding initiated the process of testing the software on individuals walking into the CSTEP/LSAMP Office. Detection of a face to initiate classification of a query image was improved by, modifying the minimum size required for an object to be recognized as a face by the cascade object detector. Upon arriving at a query image for identification, the primary task of the software, we found an identification accuracy rate of a little over 10%.
An investigation into identifying the key factors affecting the accuracy of the identification using the Histogram of Oriented Gradients (HOG) function for feature extractions and the fitcecoc function for the identification. Research suggested three factors would most likely impact the ability of the classifier to accurately identify the query image: the quantity of images for each person used to train the classifier, the variance of illumination in the images, and the deviation of the images from a full frontal image. Research on the state of facial recognition software and its various implementations together with developing a more robust database that excluded images deviating more than 45 degrees from a full frontal image, we were able to increase the success of an identification to 40%. Continued work includes improved development of the face database by standardizing the properties of the collected images and significantly increasing the size of the database.
Funder Acknowledgement(s): CNY Works Youth Services, Syracuse NY 13203; Synergy Mercy Works, Syracuse, NY, 13202; Onondaga Community College, Syracuse, NY 13215; and the Upstate Louis Stokes Alliances for Minority Participation (LSAMP) Program, Syracuse, NY 13244
Faculty Advisor: Vicentica Valdes, v.s.valdes@sunyocc.edu
Role: - Developing database of facial images - Identification and investigation of factors to minimize incorrect detection of faces - Identification of key factors affecting accuracy of image identification - Modification and coding of algorithm - Analysis