March 2015
Spotlight Summary by Kedar Khare
Thermal-to-visible face recognition using partial least squares
Automated face recognition has been an active research area for several decades, with multiple applications in biometrics and security. Typically a database of face images is stored and a suitable mathematical scheme is selected for representing distinctive features in the face images. The goal of an automated face recognition system is to determine if a given face image is close to one of the images in the database. Several research problems involving recognition in presence of variation in image resolution, face angle, illumination, pose, etc. have been studied. Most of the effort in automatic face or object recognition has been with visible light imagery. However, with increasing surveillance images being obtained from long wave infra-red (LWIR) thermal imagers operating at nighttime, correlating the faces in LWIR imagery to visible light face databases is a challenging problem. This cross-modal face recognition is a less explored area and has been addressed in detail for the first time by the present authors.
After initial pre-processing involving geometric alignment, the authors suggest the use of difference of Gaussian (DOG) filtering for reducing the modality gap between LWIR images and the visible image database. The features in this normalized image are extracted in the form of a histogram of oriented gradients (HOG) method which is dependent on the edge features in the face image. The edges in the visible image occur due to uneven reflection of light from the 3D surface structure whereas the edges in IR imagery occur due to temperature difference at the junction of various tissue types. Interestingly, a key observation that the authors make is that there is a strong correlation in the edges in the visible and IR images, and hence a good possibility of cross modal face recognition. With the features represented in the form of HOG, which may be highly correlated, the authors use a regression technique called the Partial Least Squares (PLS) for predicting the closeness of a given LWIR face image to face images in a visible image database. PLS was first developed in the field of economics but has found increasing usage in computer vision applications due to its robustness. The exhaustive study presented in this work shows the possibility of an effective thermal-to-visible face recognition system with high success rate across three face databases. The authors also present several challenges within the present framework – e.g., achieving consistent face recognition performance for face images recorded from varying distances (this is related to pixels available to represent the face) and degradation of recognition performance when the subject undergoes intense prolonged exercise causing significant changes in the thermal face image.
You must log in to add comments.
After initial pre-processing involving geometric alignment, the authors suggest the use of difference of Gaussian (DOG) filtering for reducing the modality gap between LWIR images and the visible image database. The features in this normalized image are extracted in the form of a histogram of oriented gradients (HOG) method which is dependent on the edge features in the face image. The edges in the visible image occur due to uneven reflection of light from the 3D surface structure whereas the edges in IR imagery occur due to temperature difference at the junction of various tissue types. Interestingly, a key observation that the authors make is that there is a strong correlation in the edges in the visible and IR images, and hence a good possibility of cross modal face recognition. With the features represented in the form of HOG, which may be highly correlated, the authors use a regression technique called the Partial Least Squares (PLS) for predicting the closeness of a given LWIR face image to face images in a visible image database. PLS was first developed in the field of economics but has found increasing usage in computer vision applications due to its robustness. The exhaustive study presented in this work shows the possibility of an effective thermal-to-visible face recognition system with high success rate across three face databases. The authors also present several challenges within the present framework – e.g., achieving consistent face recognition performance for face images recorded from varying distances (this is related to pixels available to represent the face) and degradation of recognition performance when the subject undergoes intense prolonged exercise causing significant changes in the thermal face image.
Add Comment
You must log in to add comments.
Article Information
Thermal-to-visible face recognition using partial least squares
Shuowen Hu, Jonghyun Choi, Alex L. Chan, and William Robson Schwartz
J. Opt. Soc. Am. A 32(3) 431-442 (2015) View: Abstract | HTML | PDF