Performance Of Generative Adversarial Networks Research Articles

Investigation the influence related to parameters configuration of Generative Adversarial Networks in face image generation

Due to the excellent performance of Generative Adversarial Networks (GAN) for age regression on face images, it is particularly important to explore the effect of different parameters on model training. In this study, the origin and development of Artificial Intelligence (AI) is first discussed, from which the concept and principles of GAN are derived. This is followed by a brief introduction of the UTKface dataset used in this research, and the Conditional Adversarial Autoencoder (CAAE) framework based on the GAN technique. The division of labor and roles of the encoder, generator, and the two discriminators in the model are described. The various learning rates as well as batch size combinations attempted in this study are then illustrated, and the training results of the model are shown in the form of graphs and plots of the loss value function. A situation where the model stops learning is highlighted in the results, which is similar to pattern descent in GAN, and is shown to be characterized by the inability of the discriminator to successfully recognize it. Ultimately, drawing from the acquired outcomes, it can be deduced that employing a larger batch size serves to enhance the pace of model training. It is advisable to concurrently elevate the learning rate by an equivalent factor when augmenting the batch size, thereby ensuring a consistent trajectory for model convergence.

Evaluating the performance of generative adversarial network-synthesized periapical images in classifying C-shaped root canals

This study evaluated the performance of generative adversarial network (GAN)-synthesized periapical images for classifying C-shaped root canals, which are challenging to diagnose because of their complex morphology. GANs have emerged as a promising technique for generating realistic images, offering a potential solution for data augmentation in scenarios with limited training datasets. Periapical images were synthesized using the StyleGAN2-ADA framework, and their quality was evaluated based on the average Frechet inception distance (FID) and the visual Turing test. The average FID was found to be 35.353 (± 4.386) for synthesized C-shaped canal images and 25.471 (± 2.779) for non C-shaped canal images. The visual Turing test conducted by two radiologists on 100 randomly selected images revealed that distinguishing between real and synthetic images was difficult. These results indicate that GAN-synthesized images exhibit satisfactory visual quality. The classification performance of the neural network, when augmented with GAN data, showed improvements compared with using real data alone, and could be advantageous in addressing data conditions with class imbalance. GAN-generated images have proven to be an effective data augmentation method, addressing the limitations of limited training data and computational resources in diagnosing dental anomalies.

Enhancing the Performance of Generative Adversarial Networks with Identity Blocks and Revised Loss Function to Improve Training Stability

Enhancing the Performance of Generative Adversarial Networks with Identity Blocks and Revised Loss Function to Improve Training Stability

Cutout with patch-loss augmentation for improving generative adversarial networks against instability

Generative adversarial networks heavily rely on large datasets and carefully chosen model parameters to avoid model overfitting or mode collapse. Cutout with patch-loss augmentation, a dataset augmentation designed for generative adversarial networks that applies cutout to both the discriminator and the generator with a patch-loss structure and a new loss function, is proposed as a solution to the issue. It can enhance the performance of generative adversarial networks on full datasets and promote better convergence and stability on limited datasets. Additionally, the tensor value clamp is proposed, accelerating training speed without compromising quality. The proposed method can be successfully used with various generative adversarial networks, according to experiments. The performance of generative adversarial networks trained with full data on CIFAR-10 is matched by our method with only 20% of the training data. Finally, combined with our approach, StyleGAN2-ADA’s Fréchet Inception Distance (FID) results on the CIFAR-10, LSUN-CAT, and FFHQ-256 datasets can be further enhanced.

A general approach for automatic segmentation of pneumonia, pulmonary nodule, and tuberculosis in CT images

Proposing a general segmentation approach for lung lesions, including pulmonary nodules, pneumonia, and tuberculosis, in CT images will improve efficiency in radiology. However, the performance of generative adversarial networks ishampered by the limited availability of annotated samples and the catastrophicforgetting of the discriminator, whereas the universality of traditional morphology-based methods is insufficient for segmenting diverse lung lesions. A cascaded dual-attention network with a context-aware pyramid feature extraction module was designed to address these challenges. A self-supervised rotation loss was designed to mitigate discriminator forgetting. The proposed model achieved Dice coefficients of 70.92, 73.55, and 68.52% on multi-center pneumonia, lung nodule, and tuberculosis test datasets, respectively. No significant decrease in accuracy was observed (p>0.10) when a small training sample size was used. The cyclic training of the discriminator was reduced with self-supervised rotation loss (p<0.01). The proposed approach is promising for segmenting multiple lung lesion types in CT images.

EEG Data Augmentation for Emotion Recognition with a Task-Driven GAN

The high cost of acquiring training data in the field of emotion recognition based on electroencephalogram (EEG) is a problem, making it difficult to establish a high-precision model from EEG signals for emotion recognition tasks. Given the outstanding performance of generative adversarial networks (GANs) in data augmentation in recent years, this paper proposes a task-driven method based on CWGAN to generate high-quality artificial data. The generated data are represented as multi-channel EEG data differential entropy feature maps, and a task network (emotion classifier) is introduced to guide the generator during the adversarial training. The evaluation results show that the proposed method can generate artificial data with clearer classifications and distributions that are more similar to the real data, resulting in obvious improvements in EEG-based emotion recognition tasks.

Assessing a Single Image in Reference-Guided Image Synthesis

Assessing the performance of Generative Adversarial Networks (GANs) has been an important topic due to its practical significance. Although several evaluation metrics have been proposed, they generally assess the quality of the whole generated image distribution. For Reference-guided Image Synthesis (RIS) tasks, i.e., rendering a source image in the style of another reference image, where assessing the quality of a single generated image is crucial, these metrics are not applicable. In this paper, we propose a general learning-based framework, Reference-guided Image Synthesis Assessment (RISA) to quantitatively evaluate the quality of a single generated image. Notably, the training of RISA does not require human annotations. In specific, the training data for RISA are acquired by the intermediate models from the training procedure in RIS, and weakly annotated by the number of models' iterations, based on the positive correlation between image quality and iterations. As this annotation is too coarse as a supervision signal, we introduce two techniques: 1) a pixel-wise interpolation scheme to refine the coarse labels, and 2) multiple binary classifiers to replace a naïve regressor. In addition, an unsupervised contrastive loss is introduced to effectively capture the style similarity between a generated image and its reference image. Empirical results on various datasets demonstrate that RISA is highly consistent with human preference and transfers well across models.

An Unsupervised Way to Understand Artifact Generating Internal Units in Generative Neural Networks

Despite significant improvements on the image generation performance of Generative Adversarial Networks (GANs), generations with low visual fidelity still have been observed. As widely used metrics for GANs focus more on the overall performance of the model, evaluation on the quality of individual generations or detection of defective generations is challenging. While recent studies try to detect featuremap units that cause artifacts and evaluate individual samples, these approaches require additional resources such as external networks or a number of training data to approximate the real data manifold. In this work, we propose the concept of local activation, and devise a metric on the local activation to detect artifact generations without additional supervision. We empirically verify that our approach can detect and correct artifact generations from GANs with various datasets. Finally, we discuss a geometrical analysis to partially reveal the relation between the proposed concept and low visual fidelity.

A novel intrinsic measure of data separability

In machine learning, the performance of a classifier depends on both the classifier model and the separability/complexity of datasets. To quantitatively measure the separability of datasets, in this study, we propose an intrinsic measure – the Distance-based Separability Index (DSI), which is independent of the classifier model. We then formally show that the DSI can indicate whether the distributions of datasets are identical for any dimensionality. DSI can measure separability of datasets because we consider the situation in which different classes of data are mixed in the same distribution to be the most difficult for classifiers to separate. And, DSI is verified to be an effective separability measure by comparing it to state-of-the-art separability/complexity measures using synthetic datasets and real datasets (CIFAR-10/100). Having demonstrated the DSI’s ability to compare distributions of samples, our other studies show that it can be used in other separability-based applications, such as measuring the performance of generative adversarial networks (GANs) and evaluating the results of clustering methods.

LiWGAN: A Light Method to Improve the Performance of Generative Adversarial Network

Generative adversarial networks (GANs) gained tremendous growth due to its potential and efficacy in producing realistic samples. This study proposes a light-weight GAN (LiWGAN) to learn the non-image synthesis with minimum computational time for less power computing. Hence, the LiWGAN method enhanced a new skip-layer channel-wise excitation module (SLE) and a self-supervised discriminator design for the non-synthesis performance using the facemask dataset. The facemask is one of the preventative strategies pioneered by the current COVID-19 pandemic. LiWGAN manipulates a non-image synthesis of facemask that could be beneficial for some researchers to identify an individual using lower power devices, occlusion challenges for face recognition, and alleviate the accuracy challenges due to limited datasets. The performance compared the processing time for a facemask dataset in terms of batch sizes and image resolutions. The Fréchet inception distance (FID) was also measured on the facemask images to evaluate the quality of the augmented image using LiWGAN. The findings for 3000 generated images showed a nearly similar FID score at 220.43 with significantly less processing time per iteration at 1.03s than StyleGAN at 219.97 FID score. One experiment was conducted using the CelebA dataset to compare with GL-GAN and DRAGAN, proving LiWGAN is appropriate for other datasets. The outcomes found LiWGAN performed better than GL-GAN and DRAGAN at 91.31 FID score with 3.50s processing time per iteration. Therefore, LiWGAN could aim to enhance the FID score to be near zero in the future with less processing time by using different datasets.

On the Performance of Generative Adversarial Network by Limiting Mode Collapse for Malware Detection Systems

Generative adversarial network (GAN) has been regarded as a promising solution to many machine learning problems, and it comprises of a generator and discriminator, determining patterns and anomalies in the input data. However, GANs have several common failure modes. Typically, a mode collapse occurs when a GAN fails to fit the set optimizations and leads to several instabilities in the generative model, diminishing the capability to generate new content regardless of the dataset. In this paper, we study conditional limiter solutions for mode collapse for the Intrusion Detection System (IDS) Control Flow GAN (ICF-GAN) model. Specifically, the ICF-GAN’s mode collapse instances are limited by a mini-batch method that significantly improves the model accuracy. Performance evaluation is conducted using numerical results obtained from experiments.

Improved Consistency Regularization for GANs

Recent work has increased the performance of Generative Adversarial Networks (GANs) by enforcing a consistency cost on the discriminator. We improve on this technique in several ways. We first show that consistency regularization can introduce artifacts into the GAN samples and explain how to fix this issue. We then propose several modifications to the consistency regularization procedure designed to improve its performance. We carry out extensive experiments quantifying the benefit of our improvements. For unconditional image synthesis on CIFAR-10 and CelebA, our modifications yield the best known FID scores on various GAN architectures. For conditional image synthesis on CIFAR-10, we improve the state-of-the-art FID score from 11.48 to 9.21. Finally, on ImageNet-2012, we apply our technique to the original BigGAN model and improve the FID from 6.66 to 5.38, which is the best score at that model size.

Generative Adversarial Network Performance in Low-Dimensional Settings.

A generative adversarial network (GAN) is an artifcial neural network with a distinctive training architecture, designed to create examples that faithfully reproduce a target distribution. GANs have recently had particular success in applications involving high-dimensional distributions in areas such as image processing. Little work has been reported for low dimensions, where properties of GANs may be better identifed and understood. We studied GAN performance in simulated low-dimensional settings, allowing us to transparently assess effects of target distribution complexity and training data sample size on GAN performance in a simple experiment. This experiment revealed two important forms of GAN error, tail underflling and bridge bias, where the latter is analogous to the tunneling observed in high-dimensional GANs.

Generative Adversarial Learning of Protein Tertiary Structures.

Protein molecules are inherently dynamic and modulate their interactions with different molecular partners by accessing different tertiary structures under physiological conditions. Elucidating such structures remains challenging. Current momentum in deep learning and the powerful performance of generative adversarial networks (GANs) in complex domains, such as computer vision, inspires us to investigate GANs on their ability to generate physically-realistic protein tertiary structures. The analysis presented here shows that several GAN models fail to capture complex, distal structural patterns present in protein tertiary structures. The study additionally reveals that mechanisms touted as effective in stabilizing the training of a GAN model are not all effective, and that performance based on loss alone may be orthogonal to performance based on the quality of generated datasets. A novel contribution in this study is the demonstration that Wasserstein GAN strikes a good balance and manages to capture both local and distal patterns, thus presenting a first step towards more powerful deep generative models for exploring a possibly very diverse set of structures supporting diverse activities of a protein molecule in the cell.

Investigating the Performance of Generative Adversarial Networks for Prostate Tissue Detection and Segmentation.

The manual delineation of region of interest (RoI) in 3D magnetic resonance imaging (MRI) of the prostate is time-consuming and subjective. Correct identification of prostate tissue is helpful to define a precise RoI to be used in CAD systems in clinical practice during diagnostic imaging, radiotherapy and monitoring the progress of disease. Conditional GAN (cGAN), cycleGAN and U-Net models and their performances were studied for the detection and segmentation of prostate tissue in 3D multi-parametric MRI scans. These models were trained and evaluated on MRI data from 40 patients with biopsy-proven prostate cancer. Due to the limited amount of available training data, three augmentation schemes were proposed to artificially increase the training samples. These models were tested on a clinical dataset annotated for this study and on a public dataset (PROMISE12). The cGAN model outperformed the U-Net and cycleGAN predictions owing to the inclusion of paired image supervision. Based on our quantitative results, cGAN gained a Dice score of 0.78 and 0.75 on the private and the PROMISE12 public datasets, respectively.

Measuring Performance of Generative Adversarial Networks on Devanagari Script

The working of neural networks following the adversarial philosophy to create a generative model is a fascinating field. Multiple papers have already explored the architectural aspect and proposed systems with potentially good results however, very few papers are available which implement it on a real-world example. Traditionally, people use the famous MNIST dataset as a Hello, World! example for implementing Generative Adversarial Networks (GAN). Instead of going the standard route of using handwritten digits, this paper uses the Devanagari script which has a more complex structure. As there is no conventional way of judging how well the generative models perform, three additional classifiers were built to judge the output of the GAN model. The following paper is an explanation of what this implementation has achieved.

Machine Learning Improves Weather and Climate Models

New research evaluates the performance of generative adversarial networks for stochastic parameterizations.

An Overview of Advances in Image Colorization Using Computer Vision and Deep Learning Techniques

Automatic image colorization as a process has been studied extensively over the past 10 years with importance given to its many applications in grayscale image colorization, aged/degraded image restoration etc. In this study, we attempt to trace and consolidate developments made in Image colorization using various computer vision techniques and methodologies, focusing on the emergence and performance of Generative Adversarial Networks (GANs). We talk in depth about GANs and CNNs, namely their structure, functionality and extent of research. Additionally, we explore the advances made in image colorization using other Deep Learning frameworks ranging from LeNets to MobileNets in order of their evolution in detail. We also compare existing published works showcasing new advancements and possibilities, and predominantly emphasize the importance of continuing research in image colorization. We further analyze and discuss potential applications and challenges of GANs to tackle in the future.

Feature-Guided SAR-to-Optical Image Translation

The powerful performance of Generative Adversarial Networks (GANs) in image-to-image translation has been well demonstrated in recent years. However, most methods are focused on completing an isolated image translation task. With the complex scenes in optical images and high-frequency speckle noise in SAR images, the quality of generated images is often unsatisfactory. In this paper, a feature-guided method for SAR-to-optical image translation is proposed to better take the unique attributes of images into account. Specifically, in view of the diversity of structure features and texture features, VGG-19 network is used as the feature extractor in the task of cross-modal image translation. To ensure the acquisition of multilayer features in the process of image generation, feature matching is carried out on different layers. Loss function based on Discrete Cosine Transform is designed to filter out the high-frequency noise. The generated images show better performance in feature preservation and noise reduction, and achieve higher Image Quality Assessment scores compared with images generated by some famous methods. The superiority of our algorithm is also demonstrated by being applied to different networks.

SAR-to-Optical Image Translation Using Supervised Cycle-Consistent Adversarial Networks

Optical remote sensing (RS) data suffer from the limitation of bad weather and cloud contamination, whereas synthetic aperture radar (SAR) can work under all weather conditions and overcome this disadvantage of optical RS data. However, due to the imaging mechanism of SAR and the speckle noise, untrained people are difficult to recognize the land cover types visually from SAR images. Inspired by the excellent image-to-image translation performance of Generative Adversarial Networks (GANs), a supervised Cycle-Consistent Adversarial Network (S-CycleGAN) was proposed to generate large optical images from the SAR images. When the optical RS data are unavailable or partly unavailable, the generated optical images can be alternative data that aid in land cover visual recognition for untrained people. The main steps of SAR-to-optical image translation were as follows. First, the large SAR image was split to small patches. Then S-CycleGAN was used to translate the SAR patches to optical image patches. Finally, the optical image patches were stitched to generate the large optical image. A paired SAR-optical image dataset which covered 32 Chinese cities was published to evaluate the proposed method. The dataset was generated from Sentinel-1 (SEN-1) SAR images and Sentinel-2 (SEN-2) multi-spectral images. S-CycleGAN was applied to two experiments, which were SAR-to-optical image translation and cloud removal, and the results showed that S-CycleGAN could keep both the land cover and structure information well, and its performance was superior to some famous image-to-image translation models.