Research On Image Retrieval Method Based On Region Of Interest

With the development of multimedia technology,the scale of large image databases is increasing,and at the same time,the demand for visual content management is also increasing.With the rapid increase in the number and scale of image databases,people are paying more and more attention to the automatic management and retrieval of digital images,which has stimulated the demand for efficient query of multimedia databases.An important solution is to use content-based image retrieval(CBIR)to directly extract features from image data and combine these features with similarity measures.In recent years,the development of deep learning has brought a new perspective to the research of image retrieval.The method of using convolutional neural networks to extract features from images has achieved good results on related data sets.However,as the complexity of image content increases and people's retrieval needs change,image retrieval gradually tends to be regionalized,sceneized,and targeted,that is,more attention is paid to the similarity of local regions of interest.Existing image retrieval research still has the problems of inaccurate localization of local interest regions,high feature dimensions,long feature processing time,low retrieval accuracy,etc.,resulting in low retrieval efficiency.In view of the above,based on the deep learning theory,this paper focuses on the analysis and application of the saliency and similarity of the image content,how to determine the location of the region of interest,extract the characteristics of the region of interest,enhance the representation of the features,and then improve the accuracy of retrieval.And to reduce the feature dimension and improve retrieval performance and other related research,the main work is as follows:Aiming at the problems of high dimension of original convolutional feature and serious occupation of computing resources,the global feature dimension obtained by using the large-scale pooling window is only 1/7 of the traditional method,which reduces the feature dimension and improves the retrieval efficiency.Aiming at the current image retrieval research trend towards scene-oriented objectification,MOP and R-MAC use the regional sampling method to obtain the randomness problem of local features,and propose an image interest region location method based on centroid theory.Using the two-dimensional global feature map of the image for centroid positioning,centering on the centroid to determine the multi-scale candidate frame and determining the position of the region of interest,compared to the sliding window sampling methods proposed by MOP and R-MAC,it can be more effective to extract local features.At the same time,the time consumed in the local feature processing stage is only 40% of R-MAC.Using Oxford5 K and Paris6 K for testing,compared with the R-MAC algorithm,the highest retrieval accuracy is improved by about 3% and 2.3%,respectively.Aiming at the problem of inaccurate localization of interest regions in traditional image retrieval methods,in order to further improve the expression quality of local features,an image retrieval method based on object detection and positioning is proposed.Use a custom object detection model to detect and locate pictures,improve the positioning accuracy of the region of interest,reduce noise interference caused by non-main objects,use local features instead of global features for feature expression,and further improve retrieval accuracy.Tested on Oxford5 K and Paris6 K,the highest retrieval accuracy reached 85.1% and 90.6% respectively,which proved the effectiveness of the proposed method.