Please use this identifier to cite or link to this item:
Title: Image aesthetic style classification and region detection using Convolutional Neural Network
Authors: Xue, Chuhui
Keywords: DRNTU::Engineering::Computer science and engineering
Issue Date: 2017
Abstract: Convolutional Neural Network (CNN) becomes popular in recent years, especially in the field of image processing. This algorithm has been successfully applied on object image classification, object detection, video analysis and so on with good results. Due to good feature extraction performance of CNN, research on automatically aesthetic analysis of images by deep learning has started. However, previous work for image aesthetic analysis like [5] are mainly about image aesthetic rating or image aesthetic binary classification. Therefore, our project aims at learning the image aesthetic styles using CNN as well as generating the bounding box of region for corresponding styles. This project comprises of two main parts, which are image aesthetic style classification and image aesthetic style region detection. We firstly build the network based on [5] and train an image aesthetic style classification model on AVA Dataset [4] with some selected style classes after data cleaning. By using this pre-trained model, we then apply Faster R-CNN [1] algorithm on image aesthetic style region detection. This is implemented by firstly manually labeling image aesthetic style region in selected images in AVA Dataset, building corresponding Region Proposal Network and Fast R-CNN Network [1] based on RAPID Network [5] and training on these labeled images with pre-trained image aesthetic style classification model.
Rights: Nanyang Technological University
Fulltext Permission: restricted
Fulltext Availability: With Fulltext
Appears in Collections:SCSE Student Reports (FYP/IA/PA/PI)

Files in This Item:
File Description SizeFormat 
  Restricted Access
3.54 MBAdobe PDFView/Open

Page view(s)

Updated on Jun 23, 2021

Download(s) 50

Updated on Jun 23, 2021

Google ScholarTM


Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.