Научная визуализация, 2022, том 14, номер 3, страницы 92 - 106, DOI: 10.26583/sv.14.3.07
Single Image DnCNN Visibility Improvement (SImDnCNNVI)
Автор: Sangita Roy1
ECE Department Narula Institute of Technology, Kolkata, India
1 ORCID: 0000-0002-8898-0183, roysangita@gmail.com
Аннотация
The presence of fog, haze, or atmospheric particles reduces visibility which is an under-constrained challenging classical problem due to ambiguous scene radiance and transmission. Consequently, digital images captured under such conditions suffer from poor recognition. In this work, a fast single image physics based inversion scattering model is adopted to overcome these limitations. Denoising convolutional neural networks (DnCNNs) model is well suited for blind Gaussian denoising in a learning framework at hidden layers. With the DnCNN blind denoised depth map, high-quality transmission is estimated and finally by inverting scattering image formation model, a clear image is obtained along with tuned haziness factor. The proposed algorithm performs well compared to sixteen state-of-the-art methods qualitatively and quantitatively on Ground Truth (GT) O-Haze dataset. Output images are appealing, halo free, edge preserved, colour balanced, clear.
Ключевые слова: Image recovery model, Quality Assessment, DnCNN, Residual Learning, Batch Normalization, Dehazing, extinction coefficient.
