DoctorateOpen Access

Noise estimation of additive white Gaussian noise with convolutional neural networks

Is this your thesis?

This record came from a bulk archive import. If it’s yours, link it to your profile.

2024
0 views
0 downloads

Abstract (EN)

Today, the usage of cameras to constitute images has an increasing importance due to increase in usage of digital technology and socialmedia. The gathered images' containing noise can not be prevented due to bothcamera's hardware propoerties and during the process of digital computations. It is highly important to remove this noise since it decreases the image quality. So, in the literature, there are lots of studies on removing the noise in images. To remove, the source of the noise should be correctly identified and modelled. Lots of non-blind denoising method, takes the noise's level from outside as parameter. In this stage, the estimation of noise level parameter plays an important role. The process of removing noise, gives successful results by determining the noise model, model parameters and lastly the level of inherent noise. By this means, noise estimation is one of the most important factors affecting the denoising success and working principle. In this study, the image gathering devices which has CMOS sensor type are defined as the application area. Real noise is modelled as Additive White Gasussian Noise (AWGN) in these devices. The image gathering devices which has CMOS sensorsa re widely used in industry, in CCTV systems, in manned/unmanned air vehicles, In this study, a novel multi frame based, independent of image and texture, a Convoutional Neural Network(CNN) noise estimation method is introduced. In this study, 4 different system model are studied. These system models are classification with 5 classes,classification with 25 classes,regression with 5 levels and regression with 25 levels. 1500 images are used for training and 500 images are used for test out of 2000 images. A success rate of %100 is achieved in 5 classs classification. In all of the models the output is the value of the standard deviation of AWGN added to the images. Since the standard deviation is the most important characteristics of the AWGN noise, this value is used in all of the models. In these models, since there is not a multiframe noisy image database in the literatüre, AWGN is added to each of the images. For all clean image two noisy images are produced. In application, this situation, being feasible with the developing technology, with rapid gathering, it means to gather images consecutively. The opposed Convolutional Neural Network (CNN) is SqueezeNet. In classification the network is used as it is whereas in regression the last two layers of the network is replaced with a Fully Connected Layer and Regression Layer. The system is also tried with ResNet and gave better results as expected. But, SqueezeNet, occupaying less storage space than others, is preferred in embedded systems. The proposed model, besides the widely used gray level images is tried on colour image databases Kodak, McMaster and BSDS500. Furthermore, the output of this system is ready to be used with other non-blind denoising methods. The output of the system is also tried on Magenetic Resonance images and the identification of image gathering devices. In regression model the system is also tried on quasi-levels of standard deviation not used while training. At the end of this try the network produced near levels of standard deviation to reel values. The system is used with non-blind denoising methods BM3D and Curvelet. The PSNR values of denoising results obtained with reel values and the estimated values are compared. This comparison shows us that the both values of the results has a linear reletionship. The system is also compared with the best results in the literatüre. The propoed method with SqueezeNet became successful 9 times out of 15, and the proposed model with ResNet became successful 15 times out of 15, when compared with the best results in the literature.

Author

Hikmet Kırmızıtaş

How to Cite

Hikmet Kırmızıtaş (Doctorate thesis). Noise estimation of additive white Gaussian noise with convolutional neural networks, 2024, Harran University.

Keywords

License

Tüm Hakları Saklıdır

This work is shared under the specified license terms.

More theses from Harran University