Research developed at Unicamp's Computing Institute (IC) offers new contributions to the method of detecting human skin in digital images. Usually, the determination is made through color analysis. In his master's thesis, supervised by professor Hélio Pedrini, computer scientist Anderson Carlos Sousa e Santos added three more parameters to the procedure: adaptive segmentation, texture and saliency. “Although it is the most important variable, color is not enough to provide accurate identification”, considers the study author.

According to Santos, the research developed at IC is inserted in a field of computing classified as “machine learning”. In the specific case of his work, the challenge was to find solutions that would contribute to refining the method for detecting human skin in digital images. The researcher explains that the exclusive analysis of color for this purpose is not effective, because of the ambiguities present in the photographs.

Examples of ambiguities are clothes, objects, backgrounds and even shadows, which often have a tone very close to that of human skin. “The intersection between these sets makes the task of distinguishing what is skin and what is not skin very difficult,” says the computer scientist. Simply put, the system developed by Santos segments the image – a kind of zoom –, initially identifying human skin through color.

Then, based on this determination, a comparison is made with the other regions of the photo, in order to decide which of them also has skin. To further refine the analysis, the system estimates the texture of the skin, which is different from the texture of clothes and objects, and also the saliency of the image, which provides an indication of the location of the skin. “By crossing these parameters, we were able to more accurately detect skin and non-skin regions”, reinforces the computer scientist.

Santos highlights that this type of research can be applied, for example, in systems used in investigations into crimes related to pedophilia. The method, he assures, has the capacity to point out, through statistical data, the presence of nudity in photographic files stored on a computer. Another possible application, adds the researcher, is in the area of ​​surveillance of uncontrolled environments.

An example is airports. By analyzing images recorded by terminal video cameras, the method is able to determine how many people are in the environment based on the amount of human skin detected. “In addition, this methodology can also be added to a biometrics system, such as the one that identifies the face. The curvature and delineation of the skin can contribute to the identification, by hypothesis, of a face contained in a database”, infers the author of the master's thesis.

Research into human skin segmentation is relatively recent in Brazil, according to the researcher. Precisely for this reason, the results obtained so far still require improvement. “My work made some contributions in this sense. Certainly, other researchers will propose new approaches with the same objective”, understands Santos, who received grants from the National Council for Scientific and Technological Development (CNPq) and the Coordination for the Improvement of Higher Education Personnel (Capes).

Now in his doctorate, the computer scientist turned his attention to another topic. He's working with video and surveillance. “It’s a different subject, but it will allow me to use the method I developed in my master’s degree. For example, this type of methodology makes it possible to confirm whether a person who was filmed by one camera is the same person filmed by another equipment. In this case, we consider information such as skin, gesture, clothing and route to make the determination”, details Santos.

Publication 

Dissertation: “Improvements in human skin segmentation in digital images”
Author Anderson Carlos Sousa and Santos
Advisor: Helio Pedrini
Unity: Institute of Computing (IC)