Image retrieval method based on contextual depth semantic information

Abstract

The invention discloses an image retrieval method based on contextual depth semantic information to mainly solve the problem that the existing image retrieval method is low in accuracy due to lack ofcontextual information. The image retrieval method based on contextual depth semantic information includes the following steps of (1) determining key points of an image by an adaptive polarization barrier method; (2) performing pre-training and fine-tuning on a convolutional neural network, constructing a polarization convolutional neural network including a region analysis layer and an iterativequantization layer; (3) extracting the contextual depth semantic features of the key points, storing the contextual depth semantic features in an index table, and completing offline indexing; (4) calculating the similarity between the image and each image in a database; (5) outputting the search result in a descending order of similarity. The image retrieval method based on contextual depth semantic information uses the contextual depth semantic features to realize the matching of the key points of the image from a region to a global environment, the proposed adaptive polarization fence methodand the constructed region analysis layer conform to the imaging characteristics of an all-sky auroral image, and the image retrieval method based on contextual depth semantic information has high retrieval accuracy and can be used for accurate retrieval of large-scale images of fisheye lens imaging.

Description

technical field [0001] The invention belongs to the technical field of image processing, relates to a deep learning algorithm and image retrieval technology, and can be used for accurate retrieval of large-scale aurora images. Background technique [0002] The high-energy charged particles carried by the solar wind hit the earth's magnetic field at high speed, sink along the "funnel"-shaped geomagnetic field lines into the north and south poles, and excite the atmospheric particles in the ionosphere to produce natural luminescence, which is the aurora. In order to study the aurora phenomenon in depth, scientists from various countries have collected massive aurora image data through platforms such as ground observation stations and space remote sensing satellites. However, affected by weather changes and cloud interference, the explosively increasing aurora image data contains a lot of invalid data. In order to study a specific solar-terrestrial space event without interfer...

AI Technical Summary

Problems solved by technology

[0005] However, there are still many deficiencies in the existing image retrieval methods: first, the existing methods only extract single-scale features of key points, lack of understanding of image context information, resulting in low retrieval accuracy; second, the existing methods use The rectangular grid method selects key points. This method does not consider the spherical distortion caused by the ground fisheye lens in the aurora image, resulting in unsatisfactory returned retrieval results.

Images