WebSAM-Adapter: Adapting Segment Anything Model for Web Page Segmentation

Bowen Ren,Zefeng Qian,Yuchen Sun,Chao Gao,Chongyang Zhang

DOI: https://doi.org/10.1007/978-3-031-56027-9_27

2024-01-01

Abstract:With the advancement of internet technology, web page segmentation, which aims to divide web pages into semantically coherent units, has become increasingly crucial for web-related applications. Conventional purely visual web page segmentation approaches, which depend on traditional edge detection, face challenges in generalizing across complex web pages. Recently, the Segment Anything Model (SAM) represents remarkable visual understanding and segmentation abilities. This inspires us that SAM can also demonstrate great potential in Web Page Segmentation. However, due to the lack of web-specific training data, its direct adaptation to web page segmentation domain has been hindered. To address this challenge, we propose WebSAM-Adapter, an effective adaptation of SAM, featuring a three-module architecture specifically tailored for web page segmentation with minimal additional trainable parameters. First, we propose a patch embedding tune module for adjusting the frozen patch embedding features, which is crucial for modifying the distribution of the original model. Second, an edge components tune module is designed to learn significant structural features within each web page. Finally, the outputs of these specialized modules are sent into our key Adapter module, which employs a lightweight multi-layer perceptron (MLP) to amalgamate these enriched features and generate webpage-specific knowledge. To the best of our knowledge, our method is the first successful adaptation of a large visual model like SAM to web page segmentation. Empirical evaluations on the comprehensive Webis-WebSeg-20 dataset demonstrate our model's state-of-the-art performance.