Image processing method and device based on neural network

Abstract

The invention discloses an image processing method and device based on a neural network to solve the problem of low image processing accuracy in the prior art. The method includes: receiving an image to be processed; using a preset neural network including at least one transformation layer as a factorized bilinear layer, processing the image to be processed to obtain a processing result; wherein, the factorized bilinear layer The output of each neuron in is the sum of the linear linear term of the input feature vector of the neuron and the factorized quadratic term representing the correlation between the input feature vectors of the neuron. In the technical solution of the present invention, the factorization quadratic item representing the correlation of the input feature vector of the neuron is added in the output result of the neuron, and the expressive ability of the neural network is improved, and the image is processed by using the neural network with strong expressive ability , which improves the accuracy of image processing.

Description

technical field [0001] The invention relates to the field of computers, in particular to an image processing method and device based on a neural network. Background technique [0002] Due to the superiority of neural networks, it is more and more common to use neural networks for image processing in the computer field. Currently, commonly used neural networks include GoogleNet, VGG, ResNet, etc. These neural networks are usually stacked by multiple basic units. The unit consists of a linear convolution layer and a nonlinear activation layer (activation functions such as tanh, sigmoid, relu, etc.), these neural networks have the ability to perform complex modeling and feature extraction on images. [0003] However, since the convolutional layers contained in the basic units of these neural networks are linear convolutional layers, the expressive ability of neural networks is limited to a certain extent. However, in practical applications, the patterns in the image are more c...

AI Technical Summary

Problems solved by technology

[0006] Scheme 2, perform a secondary outer product operation on the last convolutional layer features of the neural network to obtain the quadratic item features, and then perform averaging and final classification operations. Due to the outer product operation, the feature dimension is equal to the original feature dimension Square, this improvement scheme adds a large number of classification parameters, which increases the training complexity of the neural network and reduces the training efficiency

[0007] In summary, although there are currently proposed improvements to optimize the neural network to improve the expressiveness of the neural network, the improved solution adds a large number of parameters, which greatly increases the training complexity of the neural network and reduces the training efficiency. Difficult to apply to practical applications

Therefore, the current use of neural networks to process images still has the problem of poor accuracy.

Images