Method for extracting high-dimensional features by convolution network based on tensor

Abstract

The invention relates to a method for extracting high-dimensional features by a convolution network based on tensor. The method can be applied to classification and recognition scenes of multidimensional signals. A separation-fusion module serves a core portion of a model, and comprises a separable convolution component and a feature fusion component. The separable convolution component mainly comprises a plurality of separable convolution layers. The feature fusion component mainly comprises a multidimensional convolution core. The whole N-dimensional convolution model based on the tensor comprises a plurality of N-dimensional separation-fusion modules, pooling layers and a complete connection layer. A group of N-order tensors outputted by the last largest pooling layer are converted intoa vector to be input into the complete connection layer through vectorization, and finally, a probability vector for identification and classification is output. The method has the advantages that the network model is equivalent to existing excellent models in recognition accuracy and has much fewer parameters than the existing excellent models.

Description

technical field [0001] The invention relates to a method for extracting high-dimensional features by using a tensor-based convolution network, in particular to a convolution method based on dimensionality separability and feature fusion, which can be applied to classification and recognition scenarios of multi-dimensional signals. Background technique [0002] As a branch of machine learning, deep learning has developed very rapidly in recent years, especially in 2012, AlexNet proposed by Krizhevsky et al. has a huge advantage of 11% in the image classification competition of the large image database ImageNet. After winning the championship, the convolutional neural network has become the focus of academia again. After that, new convolutional neural network models have been proposed, such as Google's GoogLeNet, Oxford University's VGG (Visual Geometry Group), Microsoft Asia Research Institute's ResNet, etc., these networks have refreshed the record created by AlexNet on Imag...

AI Technical Summary

Problems solved by technology

[0006] However, the above network compression models are all subject to the idea of ​​model supervision, that is, first train a model with redundant parameters, and then use dimensionality reduction, decomposition and other methods to reduce model parameters on this basis, in fact, use a A lightweight model to approximate a redundant model, allowing the latter to supervise and constrain the former, needs to be trained twice. There is no problem in applying this method to two-dimensional images.

However, with the development of video coding, virtual reality and other technologies, the use of three-dimensional or even higher-dimensional image data has gradually become popular, and these methods cannot be directly applied to high-dimensional image data scenarios, because in high-dimensional Directly training a high-dimensional convolutional neural network on image data will encounter the following problems, (1) the amount of parameters to be trained is quite large; (2) the computational complexity is also very high; (3) high-dimensional samples are relatively scarce

These problems have increased the difficulty of training high-dimensional convolutional neural networks

Images