Research On Image Retrieval Technology Based On Semi-supervised Hash Learning

In the rapid development of the Internet in the 21 st century,all kinds of massive data are flooding our lives,and the speeding of daily data growth is extremely amazing,while the number of pictures in the data is the most.So image retrieval has become one of the most important technologies in people's lives.As the scale of the pictures is growing at such a high speed,how to find the information which we need effectively is necessary.If we use the images' features directly,there are some obvious disadvantages,which are the images' features having too high dimensions and taking up huge storage space.As a result,the efficiency of the retrieval process is very low.In recent years,image retrieval technology based on the hash method has been increasingly used.In traditional hash learning,using manual design for extracting images' features has limitations in retaining the semantic information of images,resulting in a poor learning effect and making the retrieval result based on hash codes worse,while the method of using deep learning to learn hash code has a relatively better effect.Most of the current hash learning using deep convolutional networks is under a supervised learning condition,but the number of tagged pictures is limited,and unlabeled data can be collected everywhere.At the same time,tagging pictures is consuming manpower and material resources.So it is particularly important to implement a image retrieval technology based on semi-supervised hash learning.In response to these problems,the work of this article is mainly as follows:Firstly,aiming to solve the problem of lack of labeled pictures,a stacked autoencoder is used to realize the image retrieval technology based on semi-supervised hash learning.Due to the performance limitation of the stacked auto-encoder,we use the generative adversarial network to improve it.Using deep network for endto-end learning can get a better performance in image retrieval tasks,because the image features extracted manually lose lots of image semantics imformation.which brings difficulties in retaining the semantic correlation between images,resulting in insufficient accuracy of image retrieval based on hash code.Then the capsule network is used to improve the step of image feature extraction,and compared with traditional deep convolution network for image featureextraction,the capsule network is more suitable due to the fact that the way of capsule network processing data is more like a brain,which can better extract the hierarchical relationships of internal representation of knowledge in the network.As a result,it has a better effect than the traditional convolutional networks for hash learning.By conducting comparative experiments in both MNIST dataset and CIFAR-10 dataset,the effect of the new network is studied to verify the effectiveness in this article.Besides compared with traditional image retrieval technology based on the hash method,the experiment shows that the semi-supervised hash learning technology implemented can achieve better result in image retrieval.