Work

A comparative study of about how image quality influences convolutional neural networks

Year: 2018
Type: book-chapter
Abstract: Deep learning in computer vision has been applied to many domains such as image classification, handwritten character classification, pedestrian detection, automatic colorization of grayscale image, content-based image retrieval, etc.While the advantages of deep architectures are widely accepted, the limitations and theoretical background are not satisfactorily researched.In this paper, we provide an evaluation of seven state-of-the-art Convolutional Neural Networks for image classification under different visual distortion types.Namely, we consider nine types of quality distortions: salt & pepper noise, median filtering, average filtering, disk filtering, periodic noise in x-and y-direction, zero-mean Gaussian noise, JPEG compression, and JPEG2000 compression.Our results indentify the distortion types that deteriorate heavily the classification performance.Furthermore, the published results may provide good funds for developing neural networks that are robust to quality distortions.
Authors: Varga, Domonkos István, Szirányi, Tamás
Language: English
Open Access status: green
OpenAlex ID: W7036950369
FWCI:
Cites: 16
Cited by:
Related to: 10