Details

Autor(en) / Beteiligte

• Le, Huu Ton
• Urruty, Thierry
• Beurton-Aimar, Marie
• Nghiem, Thi Phuong
• Tran, Hoang Tung
• Verset, Romain
• Ballere, Marie
• Lai, Hien Phuong
• Visani, Muriel

Titel

Study of CNN Based Classification for Small Specific Datasets

Ist Teil von

• Modern Approaches for Intelligent Information and Database Systems, p.355-365

Ort / Verlag

Cham: Springer International Publishing

Link zum Volltext

Quelle

Alma/SFX Local Collection

Beschreibungen/Notizen

• Recently, deep learning and particularly, Convolutional Neural Network (CNN), has become predominant in many application fields, including visual image classification. In an applicative context of detecting areas with hazard of dengue fever, we propose a classification framework using deep neural networks on a limited dataset of images showing urban sites. For this purpose, we have to face multiple research issues: (i) small number of training data; (ii) images belonging to multiple classes; (iii) non-mutually exclusive classes. Our framework overcomes those issues by combining different techniques including data augmentation and multi-scale/region-based classification, in order to extract the most discriminative information from the data. Experiment results present our framework performance using several CNN architectures with different parameter sets, without and with transfer learning. Then, we analyze the effect of data augmentation and multiscale region based classification. Finally, we show that adding a classification weighting scheme allows the global framework to obtain more than 90% average precision for our classification task.