M.Sc Thesis

M.Sc StudentMizrahi Nadav
SubjectColor-Space Classification Using Fuzzy Clustering Algorithms
DepartmentDepartment of Biomedical Engineering
Full Thesis text - in Hebrew Full thesis text - Hebrew Version


Human skin detection in color images plays an important role in various applications. This task is challenging due to the complexity of color compositions, lighting conditions and the difficulty of modeling skin features in a way that would match human interpretation. In the present research, a new improved skin detector is constructed, by incorporating information derived from applying fuzzy clustering techniques to the feature space based on the Fuzzy K-means algorithm.

The selected color space is composed of the 2 chrominance components [a,b] of the perceptually isotropic CIELab color space.

In the fuzzy k-means clustering algorithm, a correct estimation of the number of clusters and initial conditions is crucial for capturing the intrinsic structures of the feature space. To this aim, an algorithm designed to find the dominant modes of the feature space is presented. A feature space histogram is constructed in order to estimate the space density function. A pointer list - matching each bin with its local maximum is iteratively modified by simple and efficient tools until convergence is reached. This algorithm is computationally efficient and shows very good results.

The clustering carried out with a Euclidian distance fails to capture the true nature of the feature space. Also the cluster centroids are usually located close to a space mode. Hence, a “Valley metric” that detects space structures by focusing on "valleys" in the feature space is defined. These valleys are low density areas between space modes. They define the guidelines for the classification process, by checking if data vectors are positioned across a valley, with respect to a cluster centorid. Cluster assignment map obtained using the valley metric comply with feature space structures better then a map obtained using a Euclidian distance.

A simple skin detector is constructed, using a parametric threshold, with respect to the skin color model, derived from a large labeled image bank.  “Suspected” skin clusters in the feature space are identified, using the cluster assignment map derived during the classification process. Thus, an improved skin detector is constructed by applying the threshold to “suspected” skin clusters only. Doing that, an adjustment of the general skin model is made towards the skin composition of a specific image. This detector shows improved results.

Clustering methods developed in the present research are general in the sense that they consider general feature spaces, unlimited in their dimensions and structure and are thus useful in numerous applications.