Research On Image Hashing Algorithm Based On Deep Learning

With the coming of the big data era and the convenience of collection equipment,the image data spread on the Internet rapidly.The main task of image retrieval algorithms is how to retrieve the similar pictures from the massive image libraries by the given pictures.The image hashing method is a widely used image retrieval algorithm.It designs the hash function to convert the images into the hash codes,which can reduce storage space and improve image retrieval efficiency.In recent years,the rapid development of deep learning has affected all fields of computer vision.The combination of deep learning methods and image hashing algorithms greatly improves the accuracy,robustness and scalability compared to traditional hashing algorithms and it makes retrieval more efficient.This thesis explores the research of image hashing algorithms based on deep learning,aiming to improve the accuracy and efficiency of image retrieval.In both supervised and unsupervised scenarios,I innovatively design the structure and training methods of the hashing algorithms and propose solutions.The main contributions of the thesis are as follows:Firstly,in order to solve the problem of data imbalance and hard sample optimization,a supervised image hashing learning algorithm based on paired imbalance loss and disturbance improvement module is proposed.A supervised deep hashing structure for paired image input is designed.The structure uses a paired unbalanced loss function to measure the coding quality of hash codes guided by paired image tags.I use Generative Adversarial Networks to generate similar or dissimilar images,which forces the deep hashing structure to produce more robust hash codes.I set the iterative approximate quantization training method to make the codes closer to binarization during the iterative training process,thereby reducing the quantization error.Secondly,a two-stage unsupervised image hashing algorithm based on deep autoencoder is proposed for unlabeled data.The algorithm includes two training stages.The first stage of training uses the reconstructed content loss function and the binary relaxation constraint to train the autoencoder.The second stage of training uses more constraints and uses the backpropagation algorithm to iteratively update the parameters to make the generated binary code more balanced,independent and close to binarization.The feature extraction of the deep autoencoder and sufficient iterative training with constraints can ensure that the generated hash codes can get better results in each iterative training.Finally,this thesis proposes an unsupervised deep hashing algorithm based on hashing loss and Generative Adversarial Networks.The conditional Generative Adversarial Networks are combined with the unsupervised deep hashing method.The hashing loss is composed of binary relaxation constraints,coding balance loss and coding independent loss which is added to the hash coding learning.Through the image pair generation process,the hashing loss is generated for image pair loss calculation to optimize the encoder.Experiments show that the algorithm has achieved good results and the importance of key technologies in each step has also been fully verified in comparative experiments.