Appearance based pedestrians' gender recognition by employing stacked auto encoders in deep learning

Pedestrians' gender is a soft attribute which is useful in many areas of computer vision including human robot interaction, intelligent surveillance and human behavior analysis. Apart from its importance, pedestrians' gender prediction is one of the challenging methodologies in image processing. In this article, a deep learning approach is presented to classify a pedestrian as a male or a female. As a preprocessing step, pedestrian parsing is performed by a deep decompositional neural network method. The outcome of this network is a binary mask that maps the pedestrian full body from the input image. The pedestrian body image is then extracted by applying the generated pedestrian mask to the input image. This pre-processed image is then supplied to the stacked sparse auto encoder with soft max classifier for prediction. The proposed network is trained and tested separately on different pedestrians' views such as frontal views, back views and mixed views. The training is performed on PETA dataset. The experiments for testing are performed on MIT and PETA datasets (containing images other than train images). The accuracy values on MIT dataset are calculated as 82.9%, 81.8% and 82.4% on frontal, back and mixed views respectively. The mean AUC value by proposed scheme on PETA dataset is found as 91.5%+/- 4. The performance measures and comparisons with existing works depict the robustness and applicability of proposed methodology. (C) 2018 Elsevier B.V. All rights reserved.