Deep Learning Based Semantic Image Segmentation Methods for Classification of Web Page Imagery

Abstract

Semantic segmentation is the task of clustering together parts of an image that belong to the same object class. Semantic segmentation of webpages is important for inferring contextual information from the webpage. This study examines and compares deep learning methods for classifying webpages based on imagery that is obscured by semantic segmentation. Fully convolutional neural network architectures (UNet and FCN-8) with defined hyperparameters and loss functions are used to demonstrate how they can support an efficient method of this type of classification scenario in custom-prepared webpage imagery data that are labeled multi-class and semantically segmented masks using HTML elements such as paragraph text, images, logos, and menus. Using the proposed Seg-UNet model achieved the best accuracy of 95%. A comparison with various optimizer functions demonstrates the overall efficacy of the proposed semantic segmentation approach.