Enhancing Artwork Recognition through Fine-Tuning and Contrastive Learning in Deep Neural Networks
Enhancing Artwork Recognition through Fine-Tuning and Contrastive Learning in Deep Neural Networks
Requirements
- M.Sc. in Machine Learning, Data Science, Computer Science, Mathematics, Telecommunications, or similar
- Knowledge of Python
- Software development skills
- Basic concepts of image processing
- Basic concepts of data science, concerning data analysis, processing and machine learning
Description
Recognizing artworks in the middle of variations in imaging conditions poses a significant challenge in image retrieval systems. This thesis proposal centers on leveraging a dataset comprising images of artworks, including paintings, sculptures, and more, to refine the capabilities of a deep learning model. The primary aim is to fine-tune the model’s performance specifically for artwork retrieval, addressing scenarios where queries may contain images of the same artwork subject to noise or alterations during acquisition. The objective is to strengthen the model’s robustness in identifying artworks despite potential alterations, such as changes in lighting, blur, orientation, occlusion, and other factors. This adaptation aims to facilitate accurate recognition even in the presence of varying conditions, ensuring reliability in retrieving specific artworks from a database. The proposed approach involves self or semi-supervised fine-tuning of the model, employing contrastive learning techniques. This process aims to enhance the model’s ability to differentiate between similar artworks while maintaining sensitivity to new artwork not initially included in the training dataset. The emphasis lies in ensuring that the model’s robustness is complemented by its capacity to retrieve diverse artworks beyond the existing database. Throughout the thesis, the activities will include: dataset preparation and curation, model fine-tuning using self or semi-supervised learning methods, implementation and experimentation with contrastive learning techniques, evaluation of the model’s performance against varying conditions and noise levels, and validation through retrieval tasks with both known and unseen artworks. This research aligns with the flourishing field of image retrieval in art, addressing the practical need for systems capable of recognizing artworks despite potential alterations in query images. By refining deep learning models through fine-tuning and contrastive learning techniques, this work strives to advance the reliability and adaptability of artwork recognition systems, catering to real-world demands in art curation, authentication, and retrieval.