Convolutional neural network based chart image classification

Abstract

Charts are frequently embedded objects in digital documents and are used to convey a clear analysis of research results or commercial data trends. These charts are created through different means and may be represented by a variety of patterns such as column charts, line charts and pie charts. Chart recognition is as important as text recognition to automatically comprehend the knowledge within digital document. Chart recognition consists on identifying the chart type and decoding its visual contents into computer understandable values. Previous work in chart image identification has relied on hand crafted features which often fails when dealing with a large amount of data that could contain significant varieties and less common char types. Hence, as a first step towards this goal, in this paper we propose to use a deep learning-based approach that automates the feature extraction step. We present an improved version of the LeNet [LeCu 89] convolutional neural network architecture for chart image classification. We derive 11 classes of visualization (Scatter Plot, Column Chart, etc.) which we use to annotate 3377 chart images. Results show the efficiency of our proposed method with 89.5 % of accuracy rate.