Convolutional Neural Network Ensemble Fine-Tuning for Extended Transfer Learning
| Autor: | Oxana Korzh, Mikel Joaristi, Edoardo Serra |
|---|---|
| Rok vydání: | 2018 |
| Předmět: |
Training set
Contextual image classification business.industry Fire detection Computer science Deep learning Cognitive neuroscience of visual object recognition Context (language use) 02 engineering and technology 010501 environmental sciences Machine learning computer.software_genre 01 natural sciences Convolutional neural network 0202 electrical engineering electronic engineering information engineering Benchmark (computing) 020201 artificial intelligence & image processing Artificial intelligence business Transfer of learning computer 0105 earth and related environmental sciences |
| Zdroj: | Big Data – BigData 2018 ISBN: 9783319943008 BigData Congress |
| DOI: | 10.1007/978-3-319-94301-5_9 |
| Popis: | Nowadays, image classification is a core task for many high impact applications such as object recognition, self-driving cars, national security (border monitoring, assault detection), safety (fire detection, distracted driving), geo-monitoring (cloud, rock and crop-disease detection). Convolutional Neural Networks(CNNs) are effective for those applications. However, they need to be trained with a huge number of examples and a consequently huge training time. Unfortunately, when the training set is not big enough and when re-train the model several times is needed, a common approach is to adopt a transfer learning procedure. Transfer learning procedures use networks already pretrained in other context and extract features from them or retrain them with a small dataset related to the specific application (fine-tuning). We propose to fine-tuning an ensemble of models combined together from multiple pretrained CNNs (AlexNet, VGG19 and GoogleNet). We test our approach on three different benchmark datasets: Yahoo! Shopping Shoe Image Content, UC Merced Land Use Dataset, and Caltech-UCSD Birds-200-2011 Dataset. Each one represents a different application. Our suggested approach always improves accuracy over the state of the art solutions and accuracy obtained by the returning of a single CNN. In the best case, we moved from accuracy of 70.5% to 93.14%. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|