Multi-source image fusion and feature extraction algorithm based on depth learning

Abstract

The invention discloses a multi-source image fusion and feature extraction algorithm based on depth learning, comprises placing each three-dimensional model in a virtual dodecahedron, arranging a virtual camera on twenty vertices of the dodecahedron, taking virtual photos of the original object from the viewpoints of three-dimensional space, and obtaining twenty views of a single object to form amulti-view model database; The multi-view model database is divided into training set, test set and verification set according to the ratio of 7: 2: 1. The loss function is redefined by using the hidden variable of view posture label, and the loss function is minimized by back propagation algorithm. After minimizing the loss function, the last layer of the neural network outputs multiple views ofa single object through a softmax cascade, and scores of the categories under the constraint of the candidate view gesture tags. The invention avoids the dependence on the space where the feature is located, and improves the accuracy of the target classification.

Description

technical field [0001] The present invention relates to the field of multi-view target classification, in particular to a multi-source image fusion and feature extraction algorithm based on deep learning. Background technique [0002] With digitization, display technology and 3D modeling [1] With the increasing maturity of technologies such as 3D models, the goal of 3D models has shown explosive growth. How to efficiently analyze and retrieve this goal has become an important research problem. [2] . while the target class [3] As an important step in 3D model retrieval, it can greatly affect the speed and accuracy of classification. Restricting the scope of retrieval by category can not only improve the efficiency of retrieval, but also improve the accuracy of retrieval. The target classification technology refers to the classification of a part of the marked target (such as: image, video, 3D model, etc. [4] ) as a training sample, it is trained by algorithms such as deep...

AI Technical Summary

Problems solved by technology

[0005] 1. The view itself cannot represent the topology of the target 3D model, and the classification of 3D models based on the view has certain limitations

[0006] 2. How to select the optimal view affects the accuracy of the final target classification to a certain extent, and there is no relatively perfect method for the selection of the optimal view.

[0007] The main challenges currently faced in the field of view-based 3D model retrieval are [10] : Most methods focus on similarity calculation and model structure representation, while ignoring the learning and optimization of features, which limits the scope of practical applications

Images