Towards automatic photo book creation from uncurated photo collections

This project involves the creation of an automated pipeline for generating a photo book from a collection of photos. The pipeline includes the development of a neural network designed to assess the aesthetic quality of images automatically, using a state-of-the-art architecture for the task of image aesthetic assessment on the used dataset. To further improve the ability of the model of discriminating among good and bad images the RF-IRF transformation has been used to update the distribution of scores of the images. Additionally, there's a clustering process to identify similar images within the collection. This process consists into the application of the DBSCAN algorithm on the embeddings of the images, obtained from the first layer of a pre-trained MobileNetV3 network, trained on image classification. Lastly, a filtering algorithm that combines these components to create the photo book was developed. Importantly, this system offers some flexibility to the end user. One notable aspect of this project is its focus on being lightweight and compatible with mobile devices, which posed a unique challenge during development. The system demonstrated its effectiveness by outperforming a baseline model with higher computational complexity when evaluating amateur photographs and also proving to be competitive when assessing expert-level images.