NOISE CHARACTERISTICS ESTIMATION IN REAL-WORLD IMAGES USING CONVOLUTIONAL NEURAL NETWORK IN MOBILE PLATFORM

• Main Authors: Ростислав Вікторович Цехмистро, Вікторія Валеріївна Абрамова, Андрій Сергійович Рубель, Михайло Леонтійович Усс, Галина Анатоліївна Проскура, Олексій Сергійович Рубель
• Format: Article
• Language: English
• Published: National Aerospace University «Kharkiv Aviation Institute» 2019-06-01
• Series: Радіоелектронні і комп'ютерні системи
• Subjects: згорткові нейронні мереж; автоматичне оцінювання характеристик завад; мобільна платформа; оцінка якості зображень; обробка реальних зображень
• Online Access: http://nti.khai.edu/ojs/index.php/reks/article/view/760

The subject of the study is the noise characteristics in real images obtained by mobile devices. The goal is to create a demo mobile application in Android platform, which realizes real-time estimation of noise characteristics in such images. Tasks: to investigate the accuracy of noise characteristics estimation by NoiseNet neural network on test images from the Tampere17 database; to conduct a preliminary study of the type, intensity and correlation characteristics of the noise in images obtained by mobile devices; to investigate the possibility of using NoiseNet to assess the noise characteristics in these images. The following results were obtained. Analyzing the noise characteristics in test images from the Tampere17 database, distorted by white Gaussian noise, it was shown that in general, the NoiseNet neural network demonstrates a rather high estimation accuracy (the relative error of evaluation does not exceed 0.2). However, for some images, in particular, highly textured, the value of relative error can be several times higher. The noise characteristics of images taken in various conditions by cameras embedded in mobile devices from various manufacturers were studied. It is shown that the noise in such images is signal-dependent and is often characterized by a high degree of spatial correlation. At the same time, the degree of spatial correlation of noise largely depends on lighting conditions of photo taking and is higher for images obtained in dim light. Since the NoiseNet neural network is not designed to work with spatially correlated noise, for its applying the images were preprocessed to eliminate the spatial correlation of noise. The ready-to-use NoiseNet neural network and the Android demo application for testing are available on the GitHub resource: https://github.com/radiuss/NoiseNet.