ChartLine: Curve Extraction from Scientific Line Charts with Spatial-Sequence Feature Pyramid Network

Abstract Line charts are very abundant in scientific literature and commercial data visualization. Automatic curve detection from line charts is a critical early step in many upstream tasks, such as data recovery, chart quality assessment, chart plagiarism detection, visual question and answer, etc. Unlike other curve detection tasks, line charts have complex and diverse backgrounds and a variety of curve styles (e.g., solid lines, dashed lines, dotted lines, etc.), which cause existing curve detection algorithms to perform poorly on this task. This paper proposes a novel chart curve detection network (ChartLine) that includes a Spatial-Sequential Attention Feature Pyramid Network (SSA-FPN) in the encoder and decoder to learn rich curve hierarchical representations and boundary features. This model contains a Spatial Sequential Fusion Module (SSF) and a Channel Multi-Attention Module (CMA), which enhance intra-class responsiveness and inter-class discriminative ability. To demonstrate the superiority and generality of the proposed method, we evaluate it on four curve datasets and compare it with state-of-the-art curve detection, edge detection, and semantic segmentation methods. Extensive experiments show that the proposed method outperforms other state-of-the-art algorithms and achieves an F-measure of 94% on the synthetic dataset.