ExtendingText-Based Duplicate BugDe tection with Image Processing: APipeline Based Approach

Sammanfattning

Duplicate bug reports are a well-known challenge in the software development industry. Over past decades many solutions have been proposed for automatic bug duplicate detection; however, most of them focus primarily on textual fields overlooking visual data like screenshots. This study investigates screenshot-based duplicate bug report detection through an industrial collaboration with Test Scouts. It builds on a previous study that introduced Bugle, a text-based duplicate management system. We compared pipelines based on OCR-extracted text, visual embeddings, and preprocessing techniques combined with a locally run Large Language Model (LLM). Following the Design Science Research Methodology, we developed and evaluated Ilmarinen, an image-processing extension for visual duplicate detection, using HitRate@k, Precision@k, and Recall@k on acurated dataset. The results showed that the pipeline based on raw visual embeddings performed worse, indicating limited information carried by visual semantics in isolation. The raw OCR-based and LLM-enhanced pipelines achieved similar performance, with the latter showing overall better results across most metrics. It was discovered that the gathered dataset reduced the expected performance gap between them and a more diverse one would betterhighlight differences between approaches. We also concluded that a more powerful LLM would likely improve the performance of the last pipeline. For future work the study suggests integrating a Vision Language Model with a new pipeline and evaluate performance differences. Index Terms—Automated Duplicate Bug Report Detection, Image Processing, Internal Tool Development, Pipeline Architecture, Local Large Language Model, Design Science Research Methodology.