China
Africa
Explore chapters and articles related to this topic
Image Formation Using Deep Convolutional Generative Adversarial Networks
Published in Vijay Kumar, Mangey Ram, Predictive Analytics, 2021
This chapter exercises deep learning methods for image generation and restoration. DCGAN, an augmentation of the generative adversarial network, trains deep learning models to encapsulate features from training distribution to form raw data from an identical distribution. It avails itself of convolutional-transpose and convolutional layers in the generator and discriminator, respectively. The generator incorporates batch-norm layers, convolutional-transpose layers and ReLU activation layers, as shown in Figure 5.1. It maps the latent space vector (z) to data space. Seeing the data as an aggregate of images in this demonstration, converting latent space vector to data space indicates originating an RGB(red, green, blue) image (3 × 64 × 64) with a size similar to the training data in the final analysis. The discriminator, a binary classification network, inputs an image and produces a scalar probability as output that categorises the input image as fake or real. The discriminator takes a 3 × 64 × 64 input image, operates it through a sequence of stridden convolution layers, batch-norm layers and LeakyReLU activation layer and uses a sigmoid activation function to induce eventual probability. Figures 5.2 and 5.3 depicts the structure of the discriminator model.
Detecting DeepFakes for future robotic systems
Published in Brij B. Gupta, Nadia Nedjah, Safety, Security, and Reliability of Robotic Systems, 2020
Eric Tjon, Melody Moh, Teng-Sheng Moh
The base architecture of a GAN is comprised of a generator and discriminator (Figure 3.1). These two parts are implemented with deep neural networks, such as a multilayered perceptron. The generator is a network that is trained to generate samples of data. The input to the generator is a noise vector from a uniform distribution. The random noise allows the model to create varied output and capture the whole distribution of the original data. The discriminator is a classification model trained to distinguish between real and generated instance of data. The input is a single sample of data. The output is a classification score between 0 and 1 representing real or fake.
Generative Adversarial Network and Its Applications
Published in S. Kanimozhi Suguna, M. Dhivya, Sara Paiva, Artificial Intelligence (AI), 2021
A. Vijayalakshmi, Deepa V. Jose
The generator is a neural network that generates new data instances whose objective is to fool the discriminator. The discriminator evaluates the instances generated by the generator for authenticity of the instances so as not to be fooled by the generator. The generator feeds as the input a random vector at each training step, thereby preventing the model from producing the same sample every single time. The discriminator receives the synthetic sample produced by the generator and outputs the probability that the input image is a real training sample or a fake sample that is generated by the generator predicting the authenticity of the input image.
Integrating DeblurGAN and CNN to improve the accuracy of motion blur X-Ray image classification
Published in Journal of Nuclear Science and Technology, 2023
Ming-Chuan Chiu, Chia-Jung Wei
In recent years, GANs have become a leading development focus in deep learning. The initial concept of GAN, proposed by Goodfellow et al. in 2014 [13] and also known as Vanilla GAN, is defined as a game between two networks: namely, the generator and the discriminator. The goal of the generator is to recognize and learn the potential distribution in a real sample and then to generate a fake sample. The discriminator is a binary classifier whose purpose is to determine whether the input data comes from real data or from data generated by the generator. Both models continuously improve their capabilities of generation and discrimination, so the learning process of both the generator and the discriminator is also a very large and a very small game problem. When the discriminator has improved its ability to a certain extent through iterative learning and can no longer identify the source of the data, it is considered that the generator has learned the distribution of the real data [29]. As such, image blur can be represented by a general non-uniform blur as shown in Equation (1), where represents a blurred image;represents an unknown blur kernel; is a sharp image; and is additive noise.
ContactGAN development – prediction of tire-pavement contact stresses using a generative and transfer learning model
Published in International Journal of Pavement Engineering, 2022
Xiuyu Liu, Angeli Jayme, Imad L. Al-Qadi
Generative adversarial networks (GANs) are some of the most state-of-the-art artificial neural networks for generating new data based on statistics in the training dataset (Yang et al. 2021). There are two key components in the GAN: the Generator and the Discriminator. Both components are neural networks, and are trained against each other in a two-player minimax game. The Generator learns to generate samples from the training data distribution. The Discriminator learns to distinguish between the generated samples and the real samples from the dataset. Within the game theory framework, the GAN model converges to a Nash equilibrium in which the Generator produces the same distribution as the training data and the Discriminator produces 0.5 for all input data (Jiang et al. 2021). In this research, a GAN-based deep learning model named ContactGAN is proposed to generate the tire-pavement contact stresses.
A lightweight and style-robust neural network for autonomous driving in end side devices
Published in Connection Science, 2023
Sheng Han, Youfang Lin, Zhihui Guo, Kai Lv
The Discriminator is to is to determine which domain the image belongs to. The Discriminator consists of several branches that learn a binary classification determining whether an image x is a real image y or a fake one . The fake image is produced by the generator G. Similar to the typical GAN system, the Discriminator helps the generator produce more realistic images through an adversarial loss. The adversarial loss is as follows: where is the output of the Discriminator D.