Detecção automática de incompatibilidades cross-browser usando aprendizado de máquina e comparação de imagens

Abstract

Context: Cross-Browser Incompatibilities (XBIs) are compatibility issues that can be observed while rendering the same web application in different browsers. Users can interact with the Web through distinct browsers, such as: Internet Explorer, Microsoft Edge, Mozilla Firefox, Opera, Google Chrome, among others. However, the increasing number of browsers, and the constant evolution of web technologies led to differences in how browsers behave and render web applications. In order to overcome this issue during the software development process, web developers must detect and fix XBIs before deploying web applications. Many of these developers rely on manual tests of every web page rendered in several configuration environments (considering multiple platforms of operational systems and versions) to detect XBIs, regardless of the efforts and costs that are required to conduct these tasks. Goal: The goal of this research is to propose a approach of Layout XBIs automatic detection based on Machine Learning, Segmentation of the DOM Tree and Screenshot Comparison. Method: To reach the goal of this research, the process of Systematic Literature Review (SLR) was firstly executed identifing the current state of art for this research topic. Afterwards, based on the acquired knowledge, the proposed approach segmented a simple web application in multiple DOM elements. The task of XBI detection was modeled as a supervised learning classification problem using the following properties to compose the features set: differences in position, size and screenshot comparison of each DOM element of a web application. Results: The proposed approach was validated in an experiment that investigated the efficacy of the classification model. The experiment used 66 web application containing 5081 DOM elements rendered in three different browsers (Google Chrome, Mozilla Firefox, Internet Explorer). The experiment reported significant accuracy results according to F-measure metric having reached 0.91. Conclusion: The results validated the proposed approach with similar effectiveness as the state of art.