Image Copy Detection Algorithm Based On Deep Learning

Digital image grows exponentially with the development of Internet and multimedia technology,making the image spread and copy become more convenient,and copyright protection has become an urgent problem.Content-based copy detection technology,which is a new research topic in digital copyright protection,is receiving increasing attention.In image copy detection,the extracted image features need to be strongly resistant to geometric attacks.However,many copy detection algorithms cannot achieve high accuracy and efficiency.Deep learning can automatically analyze the visual content of digital image and extract features which have strong expressive ability.This thesis studies the design of image copy detection system based on deep learning.Deep learning improves the accuracy of copy detection by multi-layer neural network and learns representative visual features using massive training data.This thesis studies image feature extraction and copy detection based on deep learning.A neural network learning algorithm based on contractive auto-encoder is developed and the training algorithm of deep network is studied.This algorithm realizes automatic extraction of image features and reduces the complexity of feature extraction.The experimental results show that the neural network learning algorithm based on contractive auto-encoder can effectively increase the accuracy of image copy detection to 99.9%,which is 12.7%higher than hand-crafted visual features.