Body pixel classification by neural network

Abstract

Body pixel classification is a multiclass pixel by pixel image segmentation problem that aims to classify each image pixel to its correspondent human body part. In this article we initially adopted for this problem a Multilayer Perceptron neural network (MLP) classifier using back propagation algorithm to learn network weights and biases. Then confidence intervals based on diffMax criterion are computed in order to make classification more certain. This criterion is computed by the difference between the first and second maximum value of MLP output vector. A 92 % correct classification rate was achieved after applying confidence classification. The classification result will be integrated as an input to a human posture recognition system. © Springer-Verlag Berlin Heidelberg 2012.