Image Feature Enhancement Method Based on Database Neighborhood Relationship

Abstract

The invention discloses an image feature enhancement method based on the neighborhood relationship of a database, which includes: an improved CNN feature extraction method based on removing the size limit of the input image, and adaptively selects the optimal scale of the input image through the maximum energy pooling method, Thus, the CNN feature extraction of the database image is completed; each CNN feature in the database image is moved towards the CNN feature of its related image to a certain extent, so that the related images in the database are aggregated within a certain range in the feature space, thus Realize the enhancement of CNN features of database images. The scheme disclosed by the invention can accurately and efficiently realize the calculation of the correlation between images, so as to be applied to an image retrieval system.

Description

technical field [0001] The invention relates to the technical field of visual retrieval, in particular to an image feature enhancement method based on database neighborhood relations. Background technique [0002] In the field of image retrieval, the two most important basic problems are visual feature extraction and image similarity calculation. In the traditional image retrieval framework, most methods assume that all images are independent of each other, thus directly using the similarity between image features to measure the similarity between images. Therefore, the performance of image retrieval in this framework is highly dependent on good visual feature design. In addition to the widely used bag-of-words model based on local visual features, in recent years, there have also been many works that use convolutional neural networks (CNN) to extract global visual features output by fully connected layers for image retrieval, and have achieved good results. . However, su...

AI Technical Summary

Problems solved by technology

Both methods have achieved certain results in improving the retrieval performance. However, query expansion only uses the relationship between the database images most relevant to the query image, and does not take into account the relationship between database images; the distance measure correction requires Parameters needed to additionally save the correction measure for each image

Therefore, there is still a large room for improvement in the existing methods of using database image correlation to improve retrieval performance.

Images