A survery on CNN transfer learning for image classification

• Main Author: Teo, Jia Sheng
• Other Authors: Smitha Kavallur Pisharath Gopi
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/165184
• Institution: Nanyang Technological University

There are various ways a user can go about selecting a Convolutional Neural Net- work model for their work. The user could either self-define a model or use a pre- trained model using Transfer Learning. This work compares the two different ap- proaches and analyzes both approach. The author has achieved high accuracy val- ues with both approaches and identified the main draw back of self-defined models is the amount of computational resources it requires training a model from scratch. In addition, the author has made a comparison with 15 pre-trained Transfer Learn- ing models. From the comparisons done on the small food dataset and a larger sized Caltech101 dataset, the author concluded that the dataset is one of the main deter- minants of a model’s performance, and deeper CNN model architectures does not necessarily guarantee better results.