Skin Detection Based On A Multispectral Imaging System

With the development of computer science and digital image processing technology, object detection and recognition algorithms have attracted much attention from researchers. Skin detection is an important research topic in the application of human detection, face recognition, posture estimation, etc. Unlike skin detection method based on a traditional RGB imaging system, this thesis explores some issues related to skin detection based on a multispectral imaging system. Efforts have been made in the following aspects:1. Based on the special physiological characteristics of skin cells, the unique spectral properties of human skin in400nm-1000nm are studied in this thesis. By using spectrometer and other advanced measuring equipments, multispectral reflectance of human skin and other common object are obtained.2. A multispectral imaging system and related software platforms have been developed. Based on this system, a multispectral skin database composed by human skin images(positive samples) and other common objects(negative samples) is established.3. To simplify the multispectral imaging system and improve the efficiency of algorithms, it is better to reduce the number of spectral bands in the system. Based on the traditional sequential forward selection algorithm, an improved band selection method is proposed in the thesis. This improved algorithm uses the conditional probability to strengthen the relationship among different bands and reduces the number of used samples in each iteration. The computation load can be greatly reduced. Based on the proposed algorithm, the best band combination has been selected for multispectral skin detection.4. A skin detection algorithm based on the designed multispectral imaging system is developed. The effectiveness of the band selection results are verified by comparing the skin detection results of the selected bands with the simulated RGB bands. The multispectral reflectance of all the samples in the selected band are extracted as the input of an SVM classifier.This thesis shows that the multispectral skin detection methods can effectively detect human skin regions. Moreover, it can distinguish human skin and other similar objects, such as dummy mannequins, Silica gel masks, photos, etc.. Therefore, skin detection based on multispectral imaging systems may have great potential applications in the future.