Analysis of basic-segnet architecture with variations in training options

Abstract

Deep learning techniques are becoming popular for vision-based automation applications. Recently, various deep convolutional neural network architectures have been evolved for image classification, object detection and semantic image segmentation tasks. SegNet is one of the successful encoder-decoder convolution architectures, implemented for semantic image segmentation. We simulated Basic-SegNet architecture using MATLAB R2017b. SegNet architecture is built layer by layer without using any pre-trained model. Multi-class images from Pattern Analysis, Statistical Modelling and Computational Learning Visual Object Classes 2012 database, are used to train the architecture. The segmentation results obtained on test images are evaluated by calculating accuracy, intersection of union, boundary F1 measure and execution time. These evaluation parameters are computed over database as well as for individual object class. Training options like learning rate and its schedule, filter size, number of filters and number of epochs are varied to analyze their effects on the performance of architecture. This research work is focused on analysis of Basic-SegNet architecture with variations in training options.