Convolutional neural network image classification method based on proofreading network

Abstract

The invention discloses a convolutional neural network image classification method based on a proofreading network. A deconvolutional operation network is built on the basis of a traditional single-structure convolutional neural network to serve as the proofreading network; a first deconvolutional layer of the proofreading network is connected with a proofreading layer of a working network; a proofreading vector output layer of the proofreading network is connected with an eigenvalue tail layer of the working network through a comparison function; the proofreading layer is any convolutional layer or a full-connection layer of the working network; the eigenvalue tail layer is a layer for outputting a classification vector or a classification result; and a proofreading result is used for adjusting parameters of each network layer of the working network in a reverse sequence from back to front from the eigenvalue tail layer. The method has the beneficial effects that image data in an operation process of the convolutional neural network is restored through deconvolution; a restoration result is compared with a processing result of the convolutional neural network; and if an error occurs, the error is fed back to the original convolutional neural network, so that better training and learning can be achieved, and the accuracy of image processing is improved.

Description

technical field [0001] The present invention relates to the application of convolutional neural network in image classification, in particular, relates to a proofreading network-based convolutional neural network image classification method. Background technique [0002] Since Hinton proposed the concept of deep learning in 2006, the convolutional neural network has developed rapidly as an important direction in deep learning technology. At present, the convolutional neural network occupies a dominant position in the field of image recognition. The typical convolutional neural network model There are LeNet, VGG, ResNet, DenseNet, etc. These convolutional neural networks usually use the cross-entropy of output and input as the loss function to train the model. Although the cross-entropy of output and input can well promote the convolutional neural network model Convergence, only using cross-entropy cannot train a model with high recognition accuracy. At present, the recogniti...

AI Technical Summary

Problems solved by technology

[0002] Since Hinton proposed the concept of deep learning in 2006, the convolutional neural network has developed rapidly as an important direction in deep learning technology. At present, the convolutional neural network occupies a dominant position in the field of image recognition. The typical convolutional neural network model There are LeNet, VGG, ResNet, DenseNet, etc. These convolutional neural networks usually use the cross-entropy of output and input as the loss function to train the model. Although the cross-entropy of output and input can well promote the convolutional neural network model Convergence, only using cross-entropy cannot train a model with high recognition accuracy. At present, the recognition accuracy of convolutional neural networks in the field of medical imaging needs to be improved.

Images