Abstract:
Image processing tools to detect human skin in visible band imagery have been well
explored by many organizations, and approaches have been developed for many security
and military applications. Visible cameras are limited to human skin detection during
daylight or artificial illumination conditions, but the challenge of human skin detection
during nighttime remains an ongoing research effort. The most challenging problems are
to understand skin texture and to develop mathematical tools for discriminating skin
texture from non-skin textures in images collected using a single thermal band. To solve
this problem, a set of image processing algorithms have been designed and developed for
generating the skin-texture feature set discriminating feature selection, and classification.
First, Gray Level Co-occurrence Matrix (GLCM)-driven skin-texture features are
generated based on the skin portions of the imagery. Principal Component Analysis
(PCA) is then performed on the feature set to isolate the skin discriminating features.
Then, PCA-reported skin discriminating features are employed to construct a fused
image. The purpose of this fused image is to represent the skin pixels in terms of the
skin-discriminating features and use this image for skin discrimination. In the last
process, this fused image is used for skin and non-skin classification at the local level.
For classification, three image processing approaches are adopted: 1) Adaptive optimized
threshold with Least Mean Square algorithm, 2) Principal Component Analysis (PCA),
and 3) Linear Discriminant Analysis (LDA). Results from all three classification
techniques are analyzed for accuracy confidence levels. This research provides a
generalized approach for human skin detection in thermal images, providing a noncontact,
remote, and passive method for human skin detection in day or night imagery for
security and military applications.
