An image sample upsampling method based on convolutional self-coding

Abstract

The invention discloses an image sample upsampling method based on convolution self-coding, and the method comprises the steps of carrying out the cutting of each three-dimensional magnetic resonanceimaging sample, obtaining two-dimensional images of a region where a tumor is located through cutting, and carrying out the scale normalization of all the two-dimensional images; building a network structure in a form of cascade connection of an encoder and a decoder, and serving as a model; training the model by setting a learning rate and a loss function; carrying out optimization processing onthe trained model by adopting an adaptive moment estimation optimizer; inputting any random positive sample into the trained network to obtain low-dimensional features extracted by the encoder, calculating Euclidean distance center points of eight groups of features, and randomly selecting one group of features from the eight groups of features to obtain new features; and inputting the new features into a decoder for image reconstruction, and outputting a positive sample image. According to the method, the feature extraction is carried out through the encoder, sample enhancement is carried outon samples at the feature level, image reconstruction is carried out through the decoder, upsampling of a few types of samples is obtained, and the method can be used for balance preprocessing of classification problems.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to an image sample upsampling method based on convolutional self-encoding. Background technique [0002] Imbalanced data classification refers to the classification in which the number of samples of a certain class in the data set is much smaller than the number of other classes. The class with a large number of samples is called the majority class, and the class with a small number of samples is called the minority class [1] , in general, the classification accuracy of the majority class is much higher than that of the minority class. But in real life, minority classes often require higher accuracy, such as: disease samples in medical images, illegal intrusion samples in the Internet, and illegal cheating samples in online games. The data balance processing technology refers to normalizing the amount of different types of data to the same scale through preprocessing meth...

AI Technical Summary

Problems solved by technology

Downsampling refers to undersampling the majority class before classification, which can solve the imbalance of classification, but it will cause a waste of samples and affect the overall accuracy of classification; upsampling refers to oversampling the minority class through methods such as difference, generally However, for image data, upsampling samples directly from the image level will cause two problems:

[0004] First, the new images created may not necessarily conform to the visual senses of the human eye, for example: Synthetic minority over-sampling technique (Smote algorithm) [2] ;

[0005] Second, the new image created is easy to introduce a lot of noise, such as the upsampling algorithm based on the generated confrontation network

[0006] In response to the above problems, some methods perform upsampling from the feature level, but upsampling only by features does not have the physical meaning of iconography, and for learning algorithms such as deep learning, the relevant parameters of the feature extraction module can only be obtained through upsampling. previous samples for learning

Images