Unsupervised Classification of Multi-Class Chart Images: A Comparison of Customized CNNs and Transfer Learning Techniques

Visualization has transformed into a crucial method for presenting complex data in a reasonable design, outperforming standard representation strategies in clearness and transparency. The feasibility of growing use cases of visualization has been further enhanced by the advent Machine Learning and Deep Learning algorithms and its types such as Transfer Learning which provides the capability reuse the solutions extracted from various sources to better train the data. This paper demonstrates financial, social economic, and political data visualization charts with 4 popular classes i.e., Histogram, Bar, Line, and Pie charts. For automatic feature extraction, VGG16 model has been implemented and applying principal component analysis (PCA) to extract signal from the feature descriptors, along with k-means clustering algorithm to cluster them into 4 groups. Finally for image classification three pre trained models are used that are RESTNET50, RESTNET50 V2, DenseNet along with Customized Convolutional Neural Network on the ChartVqa dataset of graph images for image classification with following accuracy 85.40%, 80.76%, 82.46%, and 90.71% respectively. By leveraging Transfer Learning and advanced neural network models, this study advances the capability to classify and interpret chart images without the need for labeled training data.