Transfer learning based on different image retrieval models

• Main Author: Cai, Qiong
• Other Authors: Tan Yap Peng
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Electrical and electronic engineering::Computer hardware, software and systems
• Online Access: https://hdl.handle.net/10356/163989
• Institution: Nanyang Technological University

With the rapid development of digital technology, image retrieval has been used in more and more applications; for example, commodity retrieval, scenic spot retrieval, etc. However, the difficulty of collecting different types of images varies from cases to cases. Some images are easy to collect and there are large amounts of data for model training. On the other hand, some images are difficult to come by and it is not easy to have enough data for model training. Under such situation, transfer learning can be applied to address partly the problem of insufficient training set. In this dissertation, we will use two datasets (the animal dataset and the Pokemon dataset) to perform transfer learning on three image retrieval models which build on lightweight convolution network, heavyweight convolution network, and U-Net, respectively. The mAP (mean average precision) is used as the evaluation indicator to explore the effect of transfer learning for these three models. Through experiments, we conclude that the effect of model-based transfer learning on image retrieval model based on lightweight convolutional neural network is more profound. Keywords: Auto-Encoder, Convolution Neural Network, U-Net, Transfer Learning, Image Retrieval.