Wyniki wyszukiwania - BazTech

1

An analysis of denoising neural networks for noise removal in images

Ramakotti Raksha, Paneerselvam Surekha, Kumar Vinodh, Kumar Sachin, Evangeline Jebarani

Przegląd Elektrotechniczny

|

2023

|

R. 99, nr 2

27--31

EN

Clean images, when subjected to prolonged transmission, improper image acquisition or conditioned to multiple feature changes, lead to image tarnishing due to unwanted noisy pixels. This proposes to be a major threat in image-processing and computer vision fields. With the evolution of denoising models in the field of Neural Networks, efficient noise removal has become achievable, in a real-time scenario. In this work, two approaches to noise modelling have been considered, i.e., noise as an inverse problem and noise as a residual problem, this has been done by constructing convolutional auto encoders and denoising convolutional networks and their performance in the process of noise removal has been evaluated based on Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index (SSIM).

PL

Czyste obrazy poddane przedłużonej transmisji, niewłaściwej akwizycji obrazu lub poddane wielokrotnym zmianom cech prowadzą do zmatowienia obrazu z powodu niechcianych zaszumionych pikseli. Sugeruje to, że jest to poważne zagrożenie w dziedzinie przetwarzania obrazu i widzenia komputerowego. Wraz z ewolucją modeli odszumiania w dziedzinie sieci neuronowych, efektywne usuwanie hałasu stało się osiągalne w scenariuszu czasu rzeczywistego. W niniejszej pracy rozważono dwa podejścia do modelowania hałasu, tj. hałas jako problem odwrotny i hałas jako problem rezydualny. Dokonano tego poprzez skonstruowanie autoenkoderów splotowych i odszumianie sieci splotowych, a ich wydajność w procesie usuwania hałasu oceniane na podstawie stosunku sygnału szczytowego do szumu (PSNR) i wskaźnika podobieństwa strukturalnego (SSIM).

2

Mask face inpainting based on improved generative adversarial network

Liu Qingyu, Juanatas Roben A.

Applied Computer Science

|

2023

|

Vol. 19, no 2

25--42

EN

Face recognition technology has been widely used in all aspects of people's lives. However, the accuracy of face recognition is greatly reduced due to the obscuring of objects, such as masks and sunglasses. Wearing masks in public has been a crucial approach to preventing illness, especially since the Covid-19 outbreak. This poses challenges to applications such as face recognition. Therefore, the removal of masks via image inpainting has become a hot topic in the field of computer vision. Deep learning-based image inpainting techniques have taken observable results, but the restored images still have problems such as blurring and inconsistency. To address such problems, this paper proposes an improved inpainting model based on generative adversarial network: the model adds attention mechanisms to the sampling module based on pix2pix network; the residual module is improved by adding convolutional branches. The improved inpainting model can not only effectively restore faces obscured by face masks, but also realize the inpainting of randomly obscured images of human faces. To further validate the generality of the inpainting model, tests are conducted on the datasets of CelebA, Paris Street and Place2, and the experimental results show that both SSIM and PSNR have improved significantly.

3

Breast cancer detection from histopathology images using modified residual neural networks

Gupta Varun, Vasudev Megha, Doegar Amit, Sambyal Nitigya

Biocybernetics and Biomedical Engineering

|

2021

|

Vol. 41, no. 4

1272--1287

EN

Breast cancer is one of the major causes of death among women worldwide. Efficient diagnosis of breast cancer in the early phases can reduce the associated morbidity and mortality and can provide a higher probability of full recovery. Computer-aided detection systems use computer technologies to detect abnormalities in clinical images which can assist medical professionals in a faster and more accurate diagnosis. In this paper, we propose a modified residual neural network-based method for breast cancer detection using histopathology images. The proposed approach provides good performance over varying magnification factors of 40X, 100X, 200X and 400X. The network obtains an average classification accuracy of 99.75%, precision of 99.18% and recall of 99.37% on BreakHis dataset with 40X magnification factor. The proposed work outperforms the existing methods and delivers state-of-the-art results on the benchmark breast cancer dataset.

4

SpEC: A system for patient specific ECG beat classification using deep residual network

Allam Jaya Prakash, Samantray Saunak, Ari Samit

Biocybernetics and Biomedical Engineering

|

2020

|

Vol. 40, no. 4

1446--1457

EN

Electrocardiogram (ECG) is a non-invasive technique used to detect various cardiac disorders. One of the major causes of cardiac arrest is an arrhythmia. Furthermore, ECG beat classification is essential to detect life-threatening cardiac arrhythmias. The major limitations of the traditional ECG beat classification systems are the requirement of an extensive training dataset to train the model and inconsistent performance for the detection of ventricular and supraventricular ectopic (V and S) beats. To overcome these limitations, a system denoted as SpEC is proposed in this work based on Stockwell transform (ST) and two-dimensional residual network (2D-ResNet) for improvement of ECG beat classification technique with a limited amount of training data. ST, which is used to represent the ECG signal into a time-frequency domain, provides frequency invariant amplitude response and dynamic resolution. The resultant ST images are applied as input to the proposed 2D-ResNet to classify five different types of ECG beats in a patient-specific way as recommended by the Association for the Advancement of Medical Instrumentation (AAMI). The proposed SpEC system achieved an overall accuracy (Acc) of 99.73%, sensitivity (Sen) = 98.84%, Specificity (Spe) = 99.50%, Positive predictivity (Ppr) = 98.20% on MIT-BIH arrhythmia database, and shows an overall Acc of 89.87% on real-time acquired ECG dataset with classification time of single ECG beat image = 0.2365 (s) in detecting of five arrhythmia classes. The proposed method shows better performance on both the database compared to the earlier reported state-of-art techniques.