Classification, Tagging, and Object Detection in Indian Folk Paintings

Abstract

Indian folk paintings are characterized by a rich mosaic of symbols, colors, textures, and stories. These serve as invaluable repositories of cultural legacy. This thesis presents a comprehensive approach to classifying these paintings into distinct art forms and tagging them with their unique salient features. Two datasets namely FolkTalent and WarliScan, are presented. 2279 digital photos of twelve distinct Indian folk art forms make up the FolkTalent dataset. Photos are taken from websites that are direct online sellers of these artworks. GPT-4 generated followed by an expert-reviewed tags, including a broad spectrum of characteristics like color, theme, artistic style, and patterns, are appended to each artwork. On fined-tuned Convolutional Neural Network (CNN) models, classification is carried out with a remarkable accuracy of 91.83% using the RandomForest ensemble method. Deeper insights into the paintings and improved search experiences depending on thematic and visual characteristics are made possible by tagging through well-adjusted CNN-based backbones with a proprietary classifier for multi-label picture classification. Furthermore, WarliScan is presented, a unique dataset consisting of 250 digital scans of Warli paint- ings, each labeled and with exact coordinates for each unique object shown in the artwork. Originating in an Indian folk art style, Warli paintings are cultural guides as well as artistic expressions that narrate basic stories about the Warli community. WarliScan was created in order to help construct models in future for automatically verifying the authenticity of these artworks because there were no corpora with comprehensive annotations. A Mean Average Precision score of 0.585 was obtained by optimizing the YoloV8n model to create an object detection baseline that demonstrates the effectiveness of this dataset. The suggested hybrid paradigm and the combined efforts in producing the FolkTalent and WarliScan datasets established new standards for the categorization and labeling of Indian folk paintings, therefore greatly advancing the cataloging and preservation of India’s folk-art legacy.