Performance analysis of supervised classification algorithms for non parametric skin detection

Abstract

Skin segmentation is a pre-processing step in any application involving the detection of face, biometrics, gesture recognition, objectionable image blocking, human computer interaction and other recognition systems with pre-processing element being skin. The challenges in skin segmentation process are background blend, illumination variation, variation in skin tone and occlusion. The aim of this research is to analyse the performance parameters of supervised nonparametric skin classification algorithms. To achieve this, two algorithms are designed and twenty-two parameters are calculated in each case to qualitatively support the design. To address the challenges in skin detection system, Kendall Estimated Skin-Segmentation based on Heuristic Ashta Vyshistya Approach (KESHAVA) is designed. In this approach, 96 features are extracted from the input image and 12 sets are formed with initial set containing 8 features and subsequent set consisting of summation of 8 new features to the existing set. The 8 features are selected according to Kendall coefficient based on dominating performance providers. The KESHAVA algorithm is used in Classification of Human-skin Ensembles using Nonparametric necessary-feature based Neural network Algorithm (CHENNA). Supervised feature selection approach is used for classification of skin pixels and semantic segmentation network is used for training, testing and validation. Semantic segmentation network for segmenting the skin region in input images is proposed that involves minimum number of steps to arrive at the skin-segmented output. KESHAVA algorithm is tested on Pratheepan dataset and CHENNA is trained on Pratheepan dataset and tested on Pratheepan dataset, Compaq dataset, SFA dataset and Schummage dataset that are publicly available. The positive effect of the results on the qualitative and quantitative success parameters is promising