A Multi-Purpose Shallow Convolutional Neural Network for Chart Images

Abstract

Charts are often used for the graphical representation of tabular data. Due to their vast expansion in various fields, it is necessary to develop computer algorithms that can easily retrieve and process information from chart images in a helpful way. Convolutional neural networks (CNNs) have succeeded in various image processing and classification tasks. Nevertheless, the success of training neural networks in terms of result accuracy and computational requirements requires careful construction of the network layers' and networks' parameters. We propose a novel Shallow Convolutional Neural Network (SCNN) architecture for chart-type classification and image generation. We validate the proposed novel network by using it in three different models. The first use case is a traditional SCNN classifier where the model achieves average classification accuracy of 97.14%. The second use case consists of two previously introduced SCNN-based models in parallel, with the same configuration, shared weights, and parameters mirrored and updated in both models. The model achieves average classification accuracy of 100%. The third proposed use case consists of two distinct models, a generator and a discriminator, which are both trained simultaneously using an adversarial process. The generated chart images are plausible to the originals. Extensive experimental analysis end evaluation is provided for the classification task of seven chart classes. The results show that the proposed SCNN is a powerful tool for chart image classification and generation, comparable with Deep Convolutional Neural Networks (DCNNs) but with higher efficiency, reduced computational time, and space complexity.

Keywords: Siamese neural network; chart classification; convolutional neural network; data visualization; generative adversarial network; shallow neural network.

Figure 1

Architecture structure for the proposed SCNN model. The model has five weight layers.

Figure 2

Comparison of average accuracy between a proposed SCNN and a Simplified VGG. The used dataset is ICDAR 2019, with seven chart-type classes. The values are from Table 1.

Figure 3

Architecture structure for the proposed Siamese SCNN model. The model consists of two identical proposed SCNN models. The number of weight layers is doubled.

Figure 4

Comparison of average accuracy between a proposed Siamese SCNN model and a Siamese Simplified VGG model. The proposed Siamese SCNN model achieves better classification accuracy than a deeper neural network. The used dataset is ICDAR 2019, with seven chart-type classes. The values are from Table 5 and Table 6.

Figure 5

Architecture structure for the proposed DCGAN model. The model uses modified proposed SCNN model for generator and discriminator.

Figure 6

Impact of batch size and dataset size on generator and discriminator. A stable DCGAN model has a discriminator loss of around 0.5 and a generator loss of up to 4. All models are trained for 150 epochs.

Figure 7

FID and IS scores for six stable DCGAN models from Figure 6. The performance of all six models is similar, but with manual inspection, the best performing model is trained with 2500 images and batch size 32. The worst performing model is trained with 5000 images and 128 batch size.

Figure 8

The comparison of natural images and generated images. On the left is one batch with 32 natural images used in the training process. On the right side are 64 images generated by the proposed DCGAN. The used DCGAN model is trained with 2500 images and 32 batch size.

Figure 9

Impact of learning rate on DCGAN output. Multiple learning rates are tested. The proposed model is functional with any learning rate between 0.0003 and 0.0009. The best output quality is achieved with a learning rate between 0.0003 and 0.0006.