Image Denoising using Autoencoders : Denoising noisy imgaes by removing noisy pixels/grains from natural images using Deep learning and autoencoders techniques

Denoising images is widely used in applications from critical medical systems to software based image enhancement in our cell phones. The natural noise is simulated by adding noise to an image in random pixels. Currently the denoising problem can be solved in either with greedy algorithm or by deep...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:2023 IEEE 8th International Conference for Convergence in Technology (I2CT) s. 1 - 6
Hlavní autoři: Kulkarni, Uday, Patil, Sachin, E, Vikas, Patil, Rahul, Kulkarni, Bodha, M, Meena S., Shanbhag, Akshay
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 07.04.2023
Témata:
Autoencoders
Convolution
Convolution(keywords)
Deep learning
Image coding
Image resolution
MSE
Neural networks
Noise reduction
PSNR
ReLU
SGD
Sigmoid
Software
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Denoising images is widely used in applications from critical medical systems to software based image enhancement in our cell phones. The natural noise is simulated by adding noise to an image in random pixels. Currently the denoising problem can be solved in either with greedy algorithm or by deep learning techniques. In this paper we are discussing the usage of Autoencoders a deep learning technique to solve the problem of denoising. Autoencoders use down-sampling and up-sampling techniques to remove the unwanted noise from an image.By using Encoders we compress the image gradually to remove the tiny details like noise using convolution layers and ReLU as activation function. The compressed images are fed to Decoders, the decoder up-samples the image bit-by-bit till the image reaches the original resolution, while up-sampling the tiny details like human hair or tiny speckles on face are restored without trying to bring back the noisy pixels. To train the proposed model we used FFHQ-Face dataset which consists of 70 thousand images of 128x128 resolution, each image is a face of and unique individual. The output of the neural network is a denoised image with PSNR of above 35, which maximum achieved compared to other pre-trained models.
DOI:10.1109/I2CT57861.2023.10126382