Research Of Image Retrieval Based On Saliency Detection

The content-based image retrieval technology aims to find the same or similar images in the image database as output based on the input images provided by the user.This technology is now used in many fields,such as in online shopping,users can use this technology to quickly find the products they want.For another example,in the medical field,you can use this technology to search for illnesses,find parts where diseases may occur,and take precautions.Although the image retrieval task has made great progress today,the retrieval effect of the existing methods can still be improved.Most current methods use deep learning-based technology,which can extract many high-dimensional features with strong representation capabilities in the original image,but their extraction and expression of main object's information is still not good.Inadequate use main object's information in image will adversely affect search results.This article is aimed at this problem,and proposes a two-stage image retrieval model based on saliency detection.In the first stage of the model,this article first extracts the subject targets in the image,and then in the second stage performs feature extraction on the extracted image salient object.Aiming at the problem that the existing methods are not good enough in the extraction of subject information in the image and the fusion of context information in the image,this thesis proposes a saliency detection method based on multi-attention mechanism,which is the first stage of the method in this thesis.In terms of attention extraction,this method obtains the global context information by scanning in four different directions,and extracts the local context information of the image by using the full convolution network.In terms of the aggregation of context information,a method of capturing the global information and extracting the local information based on the residual structure combined with the encoder-decoder network is proposed.Finally,the effectiveness of this method is verified through comparative experiments,which shows that this method can obtain more comprehensive features and perform better saliency prediction.After the salient object is obtained,the features of the object need to be further abstracted and simplified,that is the second stage of this article.However,the current feature extraction method is not good enough in the image retrieval task to reorganize the information of the image features.In view of this problem,this thesis uses a trainable pooling method.This method can obtain parameter value that is more suitable for the task as the training process advances.Therefore,through this trainable pooling method,a more powerful image representation can be obtained from the input image.Through experiments,we can know that the image representation obtained in this way can get good results in the retrieval task,which proves the effectiveness of the method on retrieval tasks.