Improving Corruption Robustness with Random Erasing in the Frequency Domain

BackgroundDeep learning (DL) shows promise for automated lung cancer diagnosis, but limited clinical data can restrict performance. While data augmentation (DA) helps, existing methods struggle with chest computed tomography (CT) scans across diverse DL architectures.ObjectiveThis study proposes Random Pixel Swap (RPS), a novel DA technique, to enhance diagnostic performance in both convolutional neural networks and transformers for lung cancer diagnosis from CT scan images.MethodsRPS generates augmented data by randomly swapping pixels within patient CT scan images. We evaluated it on ResNet, MobileNet, Vision Transformer, and Swin Transformer models, using 2 public CT datasets (Iraq-Oncology Teaching Hospital/National Center for Cancer Diseases [IQ-OTH/NCCD] dataset and chest CT scan images dataset), and measured accuracy and area under the receiver operating characteristic curve (AUROC). Statistical significance was assessed via paired t tests.ResultsThe RPS outperformed state-of-the-art DA methods (Cutout, Random Erasing, MixUp, and CutMix), achieving 97.56% accuracy and 98.61% AUROC on the IQ-OTH/NCCD dataset and 97.78% accuracy and 99.46% AUROC on the chest CT scan images dataset. While traditional augmentation approaches (flipping and rotation) remained effective, RPS complemented them, surpassing the performance findings in prior studies and demonstrating the potential of artificial intelligence for early lung cancer detection.ConclusionsThe RPS technique enhances convolutional neural network and transformer models, enabling more accurate automated lung cancer detection from CT scan images.

Eye-tracking technology enables communication for individuals with muscle control difficulties, making it a valuable assistive tool. Traditional systems rely on electrooculography (EOG) or infrared devices, which are accurate but costly and invasive. While vision-based systems offer a more accessible alternative, they have not been extensively explored for eye-writing recognition. Additionally, the natural instability of eye movements and variations in writing styles result in inconsistent signal lengths, which reduces recognition accuracy and limits the practical use of eye-writing systems. To address these challenges, we propose a novel vision-based eye-writing recognition approach that utilizes a webcam-captured dataset. A key contribution of our approach is the introduction of a Discrete Fourier Transform (DFT)-based length normalization method that standardizes the length of each eye-writing sample while preserving essential spectral characteristics. This ensures uniformity in input lengths and improves both efficiency and robustness. Moreover, we integrate a hybrid deep learning model that combines 1D Convolutional Neural Networks (CNN) and Temporal Convolutional Networks (TCN) to jointly capture spatial and temporal features of eye-writing. To further improve model robustness, we incorporate data augmentation and initial-point normalization techniques. The proposed system was evaluated using our new webcam-captured Arabic numbers dataset and two existing benchmark datasets, with leave-one-subject-out (LOSO) cross-validation. The model achieved accuracies of 97.68% on the new dataset, 94.48% on the Japanese Katakana dataset, and 98.70% on the EOG-captured Arabic numbers dataset—outperforming existing systems. This work provides an efficient eye-writing recognition system, featuring robust preprocessing techniques, a hybrid deep learning model, and a new webcam-captured dataset.

Most Facial Expression Recognition (FER) systems rely on machine learning approaches that require large databases for an effective training. As these are not easily available, a good solution is to augment the databases with appropriate data augmentation (DA) techniques, which are typically based on either geometric transformation or oversampling augmentations (e.g., generative adversarial networks (GANs)). However, it is not always easy to understand which DA technique may be more convenient for FER systems because most state-of-the-art experiments use different settings which makes the impact of DA techniques not comparable. To advance in this respect, in this paper, we evaluate and compare the impact of using well-established DA techniques on the emotion recognition accuracy of a FER system based on the well-known VGG16 convolutional neural network (CNN). In particular, we consider both geometric transformations and GAN to increase the amount of training images. We performed cross-database evaluations: training with the "augmented" KDEF database and testing with two different databases (CK+ and ExpW). The best results were obtained combining horizontal reflection, translation and GAN, bringing an accuracy increase of approximately 30%. This outperforms alternative approaches, except for the one technique that could however rely on a quite bigger database.

A problem with convolutional neural networks (CNNs) is that they require large datasets to obtain adequate robustness; on small datasets, they are prone to overfitting. Many methods have been proposed to overcome this shortcoming with CNNs. In cases where additional samples cannot easily be collected, a common approach is to generate more data points from existing data using an augmentation technique. In image classification, many augmentation approaches utilize simple image manipulation algorithms. In this work, we propose some new methods for data augmentation based on several image transformations: the Fourier transform (FT), the Radon transform (RT), and the discrete cosine transform (DCT). These and other data augmentation methods are considered in order to quantify their effectiveness in creating ensembles of neural networks. The novelty of this research is to consider different strategies for data augmentation to generate training sets from which to train several classifiers which are combined into an ensemble. Specifically, the idea is to create an ensemble based on a kind of bagging of the training set, where each model is trained on a different training set obtained by augmenting the original training set with different approaches. We build ensembles on the data level by adding images generated by combining fourteen augmentation approaches, with three based on FT, RT, and DCT, proposed here for the first time. Pretrained ResNet50 networks are finetuned on training sets that include images derived from each augmentation method. These networks and several fusions are evaluated and compared across eleven benchmarks. Results show that building ensembles on the data level by combining different data augmentation methods produce classifiers that not only compete competitively against the state-of-the-art but often surpass the best approaches reported in the literature.

Successful training of convolutional neural networks (CNNs) requires a substantial amount of data. With small datasets, networks generalize poorly. Data Augmentation techniques improve the generalizability of neural networks by using existing training data more effectively. Standard data augmentation methods, however, produce limited plausible alternative data. Generative Adversarial Networks (GANs) have been utilized to generate new data and improve the performance of CNNs. Nevertheless, data augmentation techniques for training GANs are underexplored compared to CNNs. In this work, we propose a new GAN architecture for augmentation of chest X-rays for semi-supervised detection of pneumonia and COVID-19 using generative models. We show that the proposed GAN can be used to effectively augment data and improve classification accuracy of disease in chest X-rays for pneumonia and COVID-19. We compare our augmentation GAN model with Deep Convolutional GAN and traditional augmentation methods (rotate, zoom, etc.) on two different X-ray datasets and show our GAN-based augmentation method surpasses other augmentation methods for training a GAN in detecting anomalies in X-ray images.

Improving the prediction of extreme wind speed events with generative data augmentation techniques

This study presents an improved data augmentation technique for the classification of remote sensing (RS) scenes. The method is called Quality-based Sample Selection (QSS) data augmentation technique. It is based on generating a large number of samples using geometric transformations and then selecting the best ones based on a quality criterion. Sample images are generated online, i.e. during training of the convolutional neural network (CNN) model. For each training sample image, a few images are generated randomly using a set of geometric transformations. Then, the generated images are passed to the CNN model being trained, and the predicted probabilities are used to evaluate the quality of the images. The motivation behind this research is that we aim to augment the training set with new images that will benefit the learning of the CNN model. Therefore, an objective method to evaluate the quality of the new images is the CNN model itself. The image with the best criteria score is used to augment the training set. Several quality criteria based on the prediction probabilities of the CNN model are considered including entropy, breaking-ties and our own proposed criteria called True Label Probability (TLP). QSS is tested on five common RS scene classification datasets: UCMerced, Optimal31, RSSCN7, AID and NWPU-RS45. It has outperformed all previous methods except for one case involving NWPU-RS45 dataset. For the 10%–90% train–test split, QSS achieved 94.51, 81.72, 89.89, 94.09 and 93.98, respectively, for the mentioned datasets. While for the 20%–80% train–test split, it achieved 97.09, 88.37, 94.01, 95.71 and 94.71, respectively, for the listed datasets. Thus, this work presents a novel data augmentation method that uses data with higher quality instead of randomly selected. The paper proves experimentally that the proposed method improves the classification accuracy of RS scene classification.

This paper presents a new method for data augmentation called Stride Random Erasing Augmentation (SREA) to improve classification performance. In SREA, probability based strides of one image are pasted onto another image and also labels of both images are mixed with the same probability as the image mixing, to generate a new augmented image and augmented label. Stride augmentation overcomes limitations of the popular random erasing data augmentation method, where a random portion of an image is erased with 0 or 255 or the mean of a dataset without considering the location of the important feature(s) within the image. A variety of experiments have been performed using different network flavours and the popular datasets including fashion-MNIST, CIFAR10, CIFAR100 and STL10. The experiments showed that SREA is more generalized than both the baseline and random erasing method. Furthermore, the effect of stride size in SREA was investigated by performing experiments with different stride sizes. Random stride size showed better performance. SREA outperforms the baseline and random erasing especially on the fashion-MNIST dataset. To enable the reuse, reproduction and extension of SREA, the source code is provided in a public git repository https://github.com/kmr2017/stride-aug.

We propose Multi-scale convolutional recurrent neural networks (MCRNN) and data augmentation methods to detect polyphonic sound event with few training data. MCRNN consists of Multi-scale convolutional neural networks (MCNN) and recurrent neural networks (RNN). MCNN concatenates the higher level features extracted using multiple convolution kernels with different scales from the time domain and frequency domain at the same time. RNN is able to capture the longer term temporal context characteristics. A novel background spectrum random replacement (BSRR) data augmentation method is applied to expand training data, which uses standard normal distribution data with randomly selected position and length instead of the original time-domain, frequency-domain or time-frequency domain background spectrum features. Our method is tested on the datasets of DCASE 2019 Task3 (T3). The experimental results showed that the MCRNN and BSRR data augmentation method are efficient. We achieved better results than the first place and the single advanced on the T3 by applying BSRR and SpecAugment data augmentation method simultaneously. On the evaluation dataset (T3-eval), our best result shows 0.05 and 0.975 of error rate (ER) and F1 respectively. Our method got the best performance and relatively improved 17% and 1% than the corresponding values of the single advanced.

Diabetic retinopathy (DR) refers to a diabetes complexity that immensely impacts the eyes. This is classified into 5 various stages of the severity in accordance with the international convention. Despite that, optimization of a grading model to have a robust generalizability needs a huge number of balanced training data that is very complicated to gather, especially for greater levels of severity. A vast amount of medical data is complex and has a very high-priced method which requires cooperation between the clinics and researchers. The issue is usually attempted to be figured out with the usage of the traditional methods of data augmentation by making certain changes to images of retina dataset for instance rotation, cropping, size and zooming. In this suggested paper, the latest methods or techniques of data augmentation is exhibited which is called as deep convolutional generative adversial network (DC-GAN) and variational auto encoders (VAE). This is a particular method which is responsible for the production of artificial medical images. In addition to this, to improve DR, we can also take the aid of the classification models which are resnet50, densenet201, InceptionV3 and VGG19 for the purpose of classification of the eye related diseases. The proposed method is depicted on the Asia Pacific Tele-Ophthalmology Society (APTOS)-Blindness dataset. First, the present-day online data augmentation techniques have been utilized, and the artificial images of retina are produced by the ease of DCGAN. Then, a method of classifying is used for both techniques. Ultimately, after the method training which is done by using the real & synthetic clinical images and the outcome exhibits which the proposed model determines every stage or phase of DR and achieve the accuracy of 98.66% with using of ResNet-50 which is contrary to the current existing techniques.

One of the common challenges in medical artificial intelligence (AI) applications using echocardiography is the lack of image data harmonization. This study aims to improve the prediction accuracy of left ventricular ejection fraction (LVEF) AI models by incorporating data augmentation (DA) techniques to address the variability in image data across different vendor machines. A database comprising 15,770 echocardiographic videos from 3154 patients across five different centers was utilized, with the data acquired using various vendor machines. We prepared datasets specific to GE healthcare (GE), philips (PH), and canon medical systems (CA) vendors, including 1911, 804, and 427 cohorts, respectively. A three-dimensional convolutional neural network (3D-CNN) was trained to predict LVEF, using videos consisting of 20 images per heartbeat as input, with training performed exclusively on data from GE machines (GE-based model). DA techniques, including gamma correction, scaling, median filtering, unsharp masking, translation, rotation, noise correction, and image conversion using a cycle generative adversarial network, were applied. A regression analysis using five different chamber views was performed on the GE test data. The accuracy of LVEF prediction using the GE-based model with DA—specifically gamma correction, scaling, and translation corrections—was excellent. The mean absolute error (root mean square error) in test cohorts was 4.33 (5.58) for GE, 4.42 (5.57) for PH, and 4.89 (6.57) for CA, which were comparable to the results of a model developed using data from all vendors, demonstrating the effectiveness of DA in harmonizing images. This study demonstrates that echocardiographic videos are transferable across vendors and that DA is highly effective in improving LVEF predictions from data acquired using different vendor machines.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-21277-y.

In today's world, accurately understanding and interpreting emotions in human-computer interaction is important. In this context, this study has adopted a detailed approach to the emotion recognition problem on both speech and text data using the Interactive Emotional Dyadic Motion Capture (IEMOCAP) dataset. First, the problem of datasets with limited number of records and unbalanced distribution across classes was addressed. For this purpose, a dataset obtained from records created as improvised in the IEMOCAP dataset was used and data augmentation methods were applied for both speech and text data. Using datasets that were balanced by applying data augmentation, single-mode emotion recognition experiments were performed with models developed for Speech Emotion Recognition (SER) and Textual Emotion Recognition (TER). Subsequently, the features obtained from these two single modalities were combined with the intermediate fusion method to provide more comprehensive emotion recognition and accuracy, and the Bimodal Emotion Recognition (BiMER) system was developed. The ResNet50-CRNN+AT model, which we obtained the highest accuracy from the three different models developed for SER, creates the speech mode of BiMER, while the Bidirectional Encoder Representations from Transformers (BERT) model used for TER creates the text mode of BiMER. In this way, BiMER was supported with data augmentation methods and the robustness and generalization ability of the model were improved, reaching 88.33% accuracy. Finally, the developed BiMER system was implemented as a real-time web application using the Flask framework, and the capacity of this application to recognize emotions interactively through the user interface was tested.

Data Augmentation methods - a family of techniques designed for synthetic generation of training data - have shown remarkable results in various Deep Learning and Machine Learning tasks. Despite its widespread and successful adoption within the computer vision community, data augmentation techniques designed for natural language processing (NLP) tasks have exhibited much slower advances and limited success in achieving performance gains. As a consequence, with the exception of applications of back-translation to machine translation tasks, these techniques have not been as thoroughly explored by the wider NLP community. Recent research on the subject also still lacks a proper practical understanding of the relationship between data augmentation and several important aspects of model design, such as hyperparameters and regularization parameters. In this paper, we perform a comprehensive study of NLP data augmentation techniques, comparing their relative performance under different settings. We also propose Deep Back- Translation, a novel NLP data augmentation technique and apply it to benchmark datasets. We analyze the quality of the synthetic data generated, evaluate its performance gains and compare all of these aspects to previous existing data augmentation procedures.

Optical character recognition is a technology that turns texts and scanned documents into digital formats. The practical applications of OCR systems face a lot of challenges because of heterogeneity in scripts, font styles, and different quality images arising during their usage. The proposed work attempts to improve the robustness of OCR models built for the Hindi language by addressing the deficiencies inherent in the widely used IIIT-HW-Dev dataset. Toward this end, two methods of data augmentation are proposed: 1) synthetic images with half characters and conjuncts for complex words, and 2) various image degradations to simulate the real world. The augmented training set is used to train the two deep architectures, namely Convolutional Recurrent Neural Networks (CRNNs) and ResNet-50. Comparisons and in-depth experiments further reveal the merits of the suggested data augmentation techniques: the top performance obtained by the CRNN model exceeds current approaches. On the test set, CRNN achieved a Character Error Rate (CER) of 2.14% and Word Error Rate (WER) of 7.96%, much beyond the CER of 3.27% and WER of 11.82% achieved by the ResNet-50 model. Qualitatively, there was robustness in the case of complex words and images that were also deteriorated. The augmented dataset and trained models can be downloaded by anyone to promote more research in Hindi OCR. So, developing robust OCR systems for complex scripts starts from variability-rich and representative training data along with specific architectures, such as the CRNN architecture.

In various machine learning tasks, including image classification, data augmentation is widely used to expand the training dataset and prevent overfitting. In traditional data augmentation techniques, images from training data are flipped, distorted, added with some noise, or cropped to create new samples from the original images. This paper presents an improved and more effective data augmentation technique for image classification tasks, called Flip-and-Hide (FnH). Our technique was validated on four well-known datasets, MNIST, FMNIST, CIFAR-10 and CIFAR-100. We found 0.03% and 0.12% improvements on FMNIST and CIFAR-10, respectively, compared to the conventional Hide-and-Seek (HaS) augmentation technique. Moreover, we demonstrated the superiority of the proposed FnH augmentation technique by performing some visualisations. Consequently, our technique is particularly useful for image classification tasks that require limited training data.