Research On Image Retrieval Based On Deep Learning

Nowadays,image retrieval is widely applied.Content-based Image Retrieval(CBIR)has been widely applied in many fields,such as e-commerce,medical diagnosis,street map,copyright protection and so on.CBIR based on the content of the image(the visual feature),and retrieves the same or similar image in the existing image set.The main drawback of CBIR is that the semantic gap between low-level feature description and high-level semantics is difficult to fill,and this semantic gap cannot be eliminated.Therefore,the difficulty of improving retrieval accuracy is how to reduce the semantic gap and improve the matching degree of the feature.Deep Learning(DL)has been widely recognized in academic and industrial fields after many years of research and application.Deep learning is a machine learning method which simulates the human brain mechanism,and establishes a multi-layer neural network to learn and analyze data.It is highly efficient,universal,and malleable.Convolutional Neural Network(CNN)is one of the classic models of deep learning.Because it has the characteristics of limiting the number of parameters and mining the local structure,it is suitable for image retrieval.This thesis studies the problem of improving the performance of natural image retrieval.Instead of using the traditional low-level features of images,it uses deep learning to represent image.First,the model is trained on the classical convolutional neural network.Then the output of the full connection layer is used as the image representation vector.Finally,the similarity between the features of the search image feature database and the image to be searched for is calculated,and the statistical retrieval performance is calculated.The pooling operation not only removes useless information,but also removes useful information.Therefore,the classic CNN is improved that the pooling layer is removed to study the effect of pooling on image retrieval performance.Because the convolutional neural network only pays attention to local information,ignores the position information of the image.In order to improve this phenomenon,the attention mechanism is combined with the convolutional neural network to focus on significant regions in the image.