Automatic Classification and Feature Extraction of Bar Charts in Scanned Documents

With the introduction of high resolution cameras in mobile devices, users frequently scan document images using such devices. The presence of charts in such scanned documents are very common. Due to perceptual advantages, charts are used to represent meaningful information. Depending on the context, different chart representations are used. This paper presents a deep learning based technique for automatic classification of charts in scanned documents followed by their reconstruction using vector representation. In this paper, we presented a study of different convolution neural networks for chart classification. Based on the study, EfficientNet has been used that uniformly scales all dimensions of network depth, width and resolution using a compound coefficient for better performance. Once a chart is classified, dedicated image gradient based algorithms are used to extract chart cell, slice or bar information and these information is used for vector based reconstruction the charts. The algorithm has been applied on a number of scanned documents with charts in it. We have achieved a chart image classification accuracy of 91 % across six chart categories. Experimental results show the efficacy of the algorithm proposed.