DESCRIPTIOE
Technical field ofinvcntion:
[0002] In general, the invention relates to producing an image from a lot of data or images. To
balance the dataset
Summau of {he invention;
[0003] Conditional deep convolutional generative adversarial networks (cDCGAN),in this the
discriminative model D and_ the generative model 6 make up generative adversarial networks
(GAN). The generator model, which receives the random variable (2) as input, is a convorlutional
decoder. Using this paradigm of a noise generator, Random Variable generates random noise in
order to fabricate an image. CNN is utilized in both discriminative and generati've models to build
GAN structures. The discriminator is tricked by the generator's lifelike image simulation. A
convolutional classifier called the discriminator model can tell the difference between amhentic
and fake images. The discriminant establishes the authenticity of an image. The discriminator is in
charge of determining whether or not an image is real, while the generator creates false images in
an auempl to trick the discriminator. The two models present challenges to one another‘ The
GAN's objective is m solve the minimax issue, which can be understood as the process as a whole.
Based on log—likelihood, [he GAN (adversarial) loss function is expressed as follows: F(G, D) =
EX~Pdata (x)[log D(x)] min O max D + log(l - D(G(y)))] + EY~Py (y) First of all In this case,
py(y) is the generator's probability distribution, y is a subset of data from Py (y), x is the subset-of
data from pdata(x); G(y) is the generator network, and D(x) is the discriminator network.
Brief description ofdragving:
1. It is suggested that cDCGAN be used to create new sambles that have a distribution akin to the
original samples. A condixion vector that corresponds to a specific class or attribute and a random
noise vector are fed to the generator during training. After that, the generator creates an image
meant to correspond with the condition vector. Taking into account both the picture and the
condition vector, the discriminator is taught to differentiate between authentic images and bogus
images produced by the generator.
2. The generator and discriminator of C-DCGAN are alternately trained via the adversarial
machine learning method. Based on auxiliary (Sonditions, it can produce high-quality samples
using CNN‘s powerful feature extraction capability.
3. The model is trained and the classifier's performance is tested using the mixed samples, which
are made up of both generated and original samples, in order to assess the quality of the generated
samples. The classification ,results show that the sample data produced by the suggested CDCGAN
can be used as data augmentation to increase the classifier's capacity for generalization.
FIG 1 illustrates a framework for data augmentation method using cDCGAN
Detailed description of the invention;
[0004] Two neural networks play a game of conflict in a machine learning framework called
generative adversarial networks, or GANs. Based on the training set statistics, the two networks in
the GAN architecture will compete to generate new data in an effort to outperform the initial
results. GAN modeling, which aids in discrimination, can be enhanced with labels. Conditional
Generative adversarial networks (GAN) consist of two models: the discriminative model M and
the generative model C. These are called deep convolutional generative adversarial networks
(cDCGAN). The generator model= which receives the random variable (2) as input, is a
convolutional decoder. Using this model of a noise generator, Random Variable generates random
noise in order to fabricate an image. CNN is utilized in both discriminative and generative models to build GAN structures.
I/We claim:
1. The system claims, I wherein that claim data augmentation for brain images using cDCGAN
Sheet 1/1.