Code Localization in Programming Screencasts

摘要

Programming screencasts are growing in popularity and are often used by developers as a learning source. The source code shown in these screencasts is often not available for download or copy-pasting. Without having the code readily available, developers have to frequently pause a video to transcribe the code. This is time-consuming and reduces the effectiveness of learning from videos. Recent approaches have applied Optical Character Recognition (OCR) techniques to automatically extract source code from programming screencasts. One of their major limitations, however, is the extraction of noise such as the text information in the menu, package hierarchy, etc. due to the imprecise approximation of the code location on the screen. This leads to incorrect, unusable code. We aim to address this limitation and propose an approach to significantly improve the accuracy of code localization in programming screencasts, leading to a more precise code extraction. Our approach uses a Convolutional Neural Network to automatically predict the exact location of code in an image. We evaluated our approach on a set of frames extracted from 450 screencasts covering Java, C#, and Python programming topics. The results show that our approach is able to detect the area containing the code with 94% accuracy and that our approach significantly outperforms previous work. We also show that applying OCR on the code area identified by our approach leads to a 97% match with the ground truth on average, compared to only 31% when OCR is applied to the entire frame.