Semantic Segmentation Using ENet Architecture

摘要

Interests in augmented reality wearable's, automated devices used in home, self-driving vehicles are in high demand, which uses semantic segmentation. The methodology includes the consideration of each and every pixel, i.e., every pixel may belong to any one category and can be operated in real-time applications on less-power mobile devices. Although, huge availability of data sets and many machine learning algorithms outstrip the performances of this kind of application but as a penalty in floating-point operators which increases the running time. In order to classify the spatial images along with the meaningful segmented regions, numerous Neural Networks are proposed includes Signet or Fully CNN for multiple class classification. In this work, we used the newly introduced model ENet also called Efficient Neural Network, which is specially designed for tasks requiring low latency operations. This model is 18x times faster, using less number flops, i.e., 75 × less, and uses less number of parameter, i.e., 79 × less. The proposed method uses Cityscapes database for the experiments and the results are compared with other conventional techniques.