mohamadmomeny / create-new-data-using-denoised-images Goto Github PK

The images, corrupted by the Speckle, Gaussian, Poisson, and impulse Noise, can be restored by image enhancement approaches such as deep autoencoder networks. The pixel values in the restored data (enhanced image) and the original noise-free image are not accurately equal, depending on noise density level. Here, the dissimilarity between restored and original pixels are used as a data augmentation approach. Initially, noise of given type and density is added to the data. Next, the noise is partially eliminated from the image by employing the deep convolutional autoencoder. The denoising deep convolutional autoencoder creates the output (new data) from the noisy input, where the target is set as the original images. As a final point, the restored images are employed as new augmented data.

Deep Learning, Data augmentation, Noise, Denoised Images, Speckle Noise, Gaussian Noise, poisson Noise, Impulse Noise

Mohammad Momeny, Ali Asghar Neshat, Mohammad Arafat Hussain, Solmaz Kia, Mahmoud Marhamati, Ahmad Jahanbakhshi, Ghassan Hamarneh, "Learning-to-Augment Strategy using Noisy and Denoised Data: Improving Generalizability of Deep CNN for the Detection of COVID-19 in X-ray Images," Computers in Biology and Medicine, Volume 136, 2021, 104704, https://doi.org/10.1016/j.compbiomed.2021.104704. (https://www.sciencedirect.com/science/article/pii/S0010482521004984)

Mohammad Momeny, Ali Mohammad Latif, Mehdi Agha Sarram, Razieh Sheikhpour, Yu Dong Zhang, "A noise robust convolutional neural network for image classification, Results in Engineering," Volume 10, June 2021, 100225, https://doi.org/10.1016/j.rineng.2021.100225. (https://www.sciencedirect.com/science/article/pii/S2590123021000268)

Feel free to ask any questions. Mohammad Momeny, a Ph.D. in Computer Engineering, [email protected]