Repository logo
 

Improved Deep Learning Model Based on Integrated Convolutional Neural Networks and Transfer Learning for Shoeprint Image Classification

Files

Thesis(2.13 MB)

Date

2023

Authors

Goundar, Nikita Neha

Supervisor

Doborjeh, Maryam
Sinha, Roopak

Item type

Thesis

Degree name

Master of Information Security and Digital Forensics

Journal Title

Journal ISSN

Volume Title

Publisher

Auckland University of Technology

Abstract

Machine Learning (ML) and Deep Learning (DL) techniques have recently aided in resolving critical problems in various sectors and use cases. For example, image classification techniques based on machine learning and deep learning have proven useful in medical science and other industries. Existing work has investigated a shoeprint image classification technique to identify different classes of shoeprints for several forensic applications and wear-pattern identifications. Despite the challenges presented by the situation, there are several opportunities to explore this area. The need for more sufficient datasets is the main obstacle in this field. In addition, deep learning techniques frequently fail to achieve a high accuracy since the field still needs to be sufficiently developed. The literature has, however, used a range of traditional machine-learning algorithms. This thesis has applied deep learning classifiers for shoeprint identification. This study first explored traditional classification methods and deep learning Convolutional Neural Networks (CNN). Then it proposed a method for integrating CNN and Transfer Learning (CNN-TL) to improve the classification results. In CNN and CNN-TL methods, python's tensor-flow library was used. Finally, the shoeprint classifications were performed using a pre-trained and fine-tuned version of the Inception model, including comparing different pre-trained Inception V3, VGG16, and ResNet50 models. The results show that convolutional neural networks-transfer learning (CNN-TL) improved classification accuracy by approximately 3% compared to conventional convolutional neural networks (CNN). The study employed three techniques for shoeprint classification, namely CNN, TL, and the proposed method that combined TL with CNN with various pre-trained models (Inception V3, VGG16, and ResNet50). The performance metrics of each model employed in this study produced the following individual results: CNN model (accuracy = 96.17%), CNN-TL Inception V3 model (accuracy = 92.19%), CNN-TL VGG16 model (accuracy = 96.88%), and CNN-TL ResNet50 model (accuracy = 97.14%). The ResNet50 model achieved the highest accuracy of 97.14%, outperforming all state-of-the-art approaches in shoeprint classification. Regarding accuracy, the VGG16 model outperformed the CNN model, but the Inception V3 model performed with lower accuracy. The study highlights that the proposed methodology significantly improved the accuracy compared to previous literature. The proposed methodology is expected to open new avenues for forensic science research and deep learning approaches to image classification.

Description

Keywords

Source

DOI

Publisher's version

Rights statement

Permanent link

http://hdl.handle.net/10292/16388

Collections

Masters Theses