Intelligent 3-D Composition Recommendation System
Open Access
- Author:
- Mishra, Sahil
- Area of Honors:
- Computer Science
- Degree:
- Bachelor of Science
- Document Type:
- Thesis
- Thesis Supervisors:
- James Z Wang, Thesis Supervisor
Dr. Jesse Louis Barlow, Thesis Honors Advisor - Keywords:
- Deep learning
Recommendation System
Mobile Device
Object Detection
Object Localization
Convolutional Neural Networks - Abstract:
- As the amount of visual data around the globe increases at an exponential rate, there has been great interest among amateur and professional photographers alike to capture astonishing and aesthetically pleasing photographs that make their photographs stand out in the large crowd of digital photographs. A large percentage of the digital photographs today are taken using mobile devices. There is a lack of a recommendation system that is capable of using modern computer vision techniques to provide real-time feedback to photographers on their phones. Due to the limited hardware availability and high computational requirements of modern machine learning models, it is a difficult task to provide truly intelligent real-time recommendation systems in small devices. In this work, we explore and propose an intelligent mobile based assistant that can aid amateur photographers to take aesthetically pleasing images by altering the image composition. We propose the use of convolutional neural networks built upon the MobileNet architecture to detect and localize objects in order to understand the content and composition of the image. The inference is done on the mobile device and, hence, is faster than other methods that stream frames to a server for inference. We believe this method will help us extract multiple features from the image that can be used to assess the aesthetic quality of the image. These features also enable the use of content-based image retrieval techniques to retrieve images that are semantically similar to the ones taken by the amateur photographer.