Research On Image Retrieval Based On Deep Hash Algorithm

In recent years,with the large-scale popularization of digital cameras,smart phones and other image acquisition equipment,the innovation of image storage technology,the progress of computer vision technology and the vigorous development of the Internet,digital images show an exponential explosive growth.How to quickly and effectively retrieve the image needed by users from the massive digital image resources has been a research hotspot in the academic and industrial circles at home and abroad.At present,image retrieval technology plays a key role in trademark retrieval,medical image diagnosis,remote sensing image,public security system,digital library and other aspects of life,so the research on image retrieval technology has great practical significance and application value.The traditional text-based image retrieval technology needs manual marking features,which spends a lot of time and manpower,and this method has great limitations in the highlevel semantic representation of images.At the beginning,there is still a big semantic gap in content-based image retrieval.Researchers at home and abroad have been looking for a better model of human visual perception.With the successful application of deep learning in computer vision,object recognition and other fields,it brings good news to the field of image retrieval.More and more researchers apply deep learning to image retrieval.The convolution neural network in the deep learning can learn and train well to get the high-level semantic features of the image,and retrieve the image intelligently.Based on deep learning,this paper builds a model which can realize efficient and accurate image retrieval based on depth hash algorithm.Image retrieval technology can be roughly summarized as image feature extraction and similarity measurement technology,so we can consider the establishment of depth hash network retrieval model from these two aspects.In the process of feature extraction,based on the classical convolutional neural network ZFNet,a new layer is added in the last pooling layer,and the key is to add a hidden layer.In this paper,sigtan,a combined function of sigmoid and tanh,is proposed.By adding a hidden layer between the input layer and its upper layer,sigtan is used as the activation function of the hidden layer.The significance of introducing this function is to transform image features into corresponding binary sequences,that is hashing.By using binary sequence to match the similar image first,on the one hand,it can improve the retrieval accuracy to a certain extent,on the other hand,it can greatly reduce the scope of searching feature map,and then improve the efficiency of retrieval;in the similarity measurement,because the extracted image feature dimension is very high,it increases the complexity of calculation,and also has a certain impact on the retrieval accuracy.Therefore,the classic multidimensional scaling MDS algorithm is selected for feature dimensionality reduction.Integrating the above two links can build a deep hash retrieval model.First of all,a deep hash network retrieval model is built through theoretical research and experimental proof,and then the retrieval performance of the model is explored through experiments on the general data set.The experimental results show that the model can efficiently and accurately retrieve the images required by users,and the retrieved images are arranged in descending order according to the degree of relevance.Finally,a convenient deep hash network retrieval system is built.In conclusion,the research of image retrieval using depth hash algorithm is easy to practice and has good application value.