Spatial information learning method based on artificial neural network

Abstract

The invention relates to a spatial information learning method based on an artificial neural network, and belongs to the technical field of deep learning. The spatial information learning method includes utilizing a traditional neural network to perform abstract expression on an input picture, and extracting feature maps from medium and high layers of the traditional neural network to obtain a feature vector F; performing filtering mapping operation on F to obtain LF to enhance data significance; performing convolution operation on a convolutional neural network through LF to obtain more abstract feature expression X; performing GAP dimension reduction processing on X to obtain feature points X*; performing correlation analysis on X* through a formula D=(X*)*(X*)<T> to obtain a correlation matrix; performing projection operation on D through a formula CD=D*V to obtain a structure feature vector CD; and performing fusion operation on CD through a formula Y(F)=F+[lambda]*CD to obtain and output final features. Compared with the prior art, the spatial information learning method based on the artificial neural network can identify an object to which geometric deformation happens without depending on a data enhancement method, so that the neural network is more robust to identification of invariance.

Description

technical field [0001] The invention relates to a spatial information learning method based on an artificial neural network, which is a method for extracting invariant features by using a neural network, in particular to a method for extracting spatial information by using a deep convolutional neural network, and belongs to the technical field of deep learning. Background technique [0002] Convolutional Neural Network (CNN), as one of the best feature extractors at the present stage, not only has its expressiveness shine in the field of computer vision, but also the local perception principle of convolution kernel has made NLP, Go game (AlphaGo) There has been progress in the non-visual field. In order to extract more expressive features, more researchers are working on how to design an efficient convolutional neural network architecture. For example, Alexnet—it is the first leap-like result improvement of convolutional neural network in recent times, VGG, Inception—can ma...

AI Technical Summary

Problems solved by technology

[0006] The purpose of the present invention is to solve the problem that the artificial neural network relies on data preprocessing methods to improve the generalization ability of the model at the present stage, and proposes a spatial information learning framework, so that the features captured by the artificial neural network are invariant and reduce the number of artificial neural networks. Dependency on data preprocessing

Images