Sky-CNN: A CNN-based Learning Approach for Skyline Scene Understanding

Ameni Sassi (1) , Wael Ouarda (2) , Chokri Ben Amar , Serge Miguet (3, 4)

1 REGIM-Lab - REsearch Group in Intelligent Machines [Sfax]
2 ENIS - École Nationale d'Ingénieurs de Sfax | National School of Engineers of Sfax
3 imagine - Extraction de Caractéristiques et Identification
4 LIRIS - Laboratoire d'InfoRmatique en Image et Systèmes d'information

Skyline scenes are a scientific matter of interest for some geographers and urbanists. These scenes have not been well-handled in computer vision tasks. Understanding the context of a skyline scene could refer to approaches based on hand-crafted features combined with linear classifiers; which are somewhat side-lined in favor of the Convolutional Neural Networks based approaches. In this paper, we proposed a new CNN learning approach to categorize skyline scenes. The proposed model requires a pre-processing step enhancing the deep-learned features and the training time. To evaluate our suggested system; we constructed the SKYLINEScene database. This new DB contains 2000 images of urban and rural landscape scenes with a skyline view. In order to examine the performance of our Sky-CNN system, many fair comparisons were carried out using well-known CNN architectures and the SKYLINEScene DB for tests. Our approach shows it robustness in Skyline context understanding and outperforms the hand-crafted approaches based on global and local features.

Mots clés: Convolutional Neural Network; deep learning; scene categorization; skyline; features representation; deep learned features

Domaines: Méthodes et statistiques; Traitement des images [eess.IV]

Ameni Sassi, Wael Ouarda, Chokri Ben Amar, Serge Miguet. Sky-CNN: A CNN-based Learning Approach for Skyline Scene Understanding. International Journal of Intelligent Systems and Applications, 2019, 4, pp.14 - 25. ⟨10.5815/ijisa.2019.04.02⟩. ⟨halshs-02471883⟩