| Summary: | 碩士 === 國立清華大學 === 電機工程學系 === 101 === Varying lighting conditions affects the performance of a facial recognition system. HOG is robust feature to use under lighting variations as image brightness changes smoothly and slowly. However, under severe changes in brightness, for example boundaries of light and shade, the HOG feature becomes unstable.
This thesis proposes Face Recognition Using Dynamically Weighted HOG Based on Illumination Conditions. This method extracts Illumination Feature around the landmarks of local features, and evaluates the stability of HOG feature by Illumination Feature differences. Two methods are proposed to improve the recognition rate: (1) Amend unstable HOG feature itself. (2) Improve matching algorithm according to the overall lighting condition. The first method uses the characteristics of the HOG feature. HOG uses the gradient orient and intensity to describe facial details around landmarks, which is similar to the characteristics of directional lighting variations. Therefore, detecting the location (and direction) of lighting variation will allow amendments to HOG feature. The second method involves adding facial landmarks coordinate as Shape feature to determine the weighting of both HOG and Shape feature according to the overall lighting conditions.
Experimental results on CMU PIE database show that Illumination feature can describe the lighting conditions around landmarks successfully. On this basis, we further quantify the overall lighting conditions by the numbers and directions of landmarks under varying lighting conditions. The weighting of both HOG and Shape feature can be determined based on overall lighting conditions, and reduce the FRR from 2.8% to 1.5%. This achieves Face Recognition Using Dynamically Weighted HOG Based on Illumination Conditions.
|
|---|
