Visual search using convolutional neural networks

• Main Author: Ng, Sue Shen
• Other Authors: Yap Kim Hui
• Format: Final Year Project
• Language: English
• Published: 2018
• Subjects: DRNTU::Engineering
• Online Access: http://hdl.handle.net/10356/74976
• Institution: Nanyang Technological University

In recent years, deep learning has provided the breakthrough of many new practical applications of machine learning. One such deep learning approach is convolutional neural networks (CNNs). This study will introduce the model structure and principle of the CNN, as well as the development of one for the purpose of skin cancer image classification. The development of the CNN involved the transfer learning of the VGG-VD pre-trained model, vgg-verydeep-16, using MATLAB. To facilitate the improvement in accuracy of the skin cancer classification, controlled trainings were conducted with varied learning rates and mini-batch size, and application of data augmentation. The resultant CNN is able to classify melanoma and mole images with an accuracy of 83.0%. What differentiates the architecture of a CNN from other neural networks is the multiple layers of convolution and pooling before the fully connected layers we see in artificial neural networks (ANN). The difference in architecture has allowed CNNs to be successful in 2D image recognition and classification, hence it is used for this project.