Please use this identifier to cite or link to this item:
|Title:||Deep learning for image classification||Authors:||Ku, Wee Kiat||Keywords:||DRNTU::Engineering::Computer science and engineering||Issue Date:||2017||Abstract:||Recent advancements in Convolutional Neural Networks (CNN) has been highly successful in more tricky and complex image processing tasks like object detection and classification. Research has also been conducted on classifying painting and photographic aesthetic styles (Karayev et al., 2014) with CNNs. However, such work on classifying photographic aesthetic styles are often limited as datasets used are multi-class, single label datasets but in reality multiple aesthetic styles can co-exist together in a single photograph. Therefore, this project aims to provide a classification pipeline that is able to provide multi-label results from a multi-class, single-label dataset and to build a photographic aesthetic style tool on top of the pipeline for photographers to improve on the aesthetic styles of their photographs. This project consists of 2 main parts, photographic aesthetic style classification pipeline and aesthetic style settings recommendation system. A CNN architec- ture, AlexNet (Krizhevsky, Sutskever, & Hinton, 2012), was chosen to be trained with the AVA Dataset (Murray, Marchesotti, & Perronnin, 2012). Through a series of experiments it was determined that the AVA Dataset has a lot of mis- labelled images, a new dataset was then collected to train the CNN with. The new dataset includes photograph meta data, EXIF, which is used to train binary Random Forest classifiers as part of the classification pipeline. A camera settings recommender system was then built on top of the classification pipeline. Accessed through a web API, the system is able to classify photographs to aesthetic styles as well as recommend camera settings given a photograph or an aesthetic style chosen by the user.||URI:||http://hdl.handle.net/10356/72800||Rights:||Nanyang Technological University||Fulltext Permission:||restricted||Fulltext Availability:||With Fulltext|
|Appears in Collections:||SCSE Student Reports (FYP/IA/PA/PI)|
Updated on Nov 28, 2020
Updated on Nov 28, 2020
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.