Researc On Accurate Target Retrieval Based On Feature Expression Of Region Of Interest

In recent years,how to effectively retrieve the desired target image from the complex image content has become an urgent problem to be solved:(1)In current image retrieval models,the underlying features of the image are usually extracted for the overall image,and less information about the area of interest in the image is considered less;(2)For images with multiple targets Retrieval tasks cannot effectively perform multi-target region queries;(3)When images exist in two different domain distributions,effective cross-domain retrieval cannot be performed.In response to the above problems,this paper proposes a method for research on accurate target retrieval based on feature expression of region of interest,object detection technology is used to obtain the area of interest in the image,filtering out the interference of invalid background information,using feature fusion to obtain the target area and global effective features,and simultaneously adding domain confrontation structures to the images of different domains to achieve cross-domain retrieval.The main work of this article is as follows:1.A feature retrieval method for image instance-level region of interest is proposed.Using the method of target detection,the features of the target area are extracted as effective areas of interest,so as to filter out invalid background information in the entire image to obtain a more accurate retrieval effect.2.A method for weighted fusion of regional features is proposed.Use the class probability of the existing multi-target as the weight of the class feature,and weight the features with multiple target features to obtain instance-level features that can represent the entire image's interest area;meanwhile,use the attention mechanism to enhance the feature's sensitivity to the area degree,and fused with instance-level features of the target area of interest,so that the influence of global features and target features in the image can be taken into account,which can better meet the purpose of retrieval and improve retrieval accuracy.3.A multi-level aligned domain adaptation method is proposed for the features of the target area of interest.By adding a dual confrontation structure to the feature learning process,the existing feature confrontation method is improved,making it more suitable for the features of the target area of interest,and effective retrieval of two different domain images is achieved.Through verification analysis,the effectiveness of the retrieval framework proposed in this paper is confirmed.In the retrieval of real scenes,mAP is increased by 5.2%,and in cross-domain target acquisition,mAP is increased by 7%.At the same time,by visualizing the retrieval results,the method presented in this paper shows good performance.