Transfer learning approach for classification and noise reduction on noisy web data

One of the main ingredients to learn a visual representation of an object using the Convolutional Neural Networks is a large and carefully annotated dataset. Acquiring a dataset in a demanded scale is not a straightforward task; therefore, the community attempts to solve this problem by creating noisy datasets gathered from web sources. In this paper, this issue is tackled by designing a vehicle recognition system using Convolutional Neural Networks and noisy web data. In the proposed system, the transfer learning technique is employed, and behavior of several deep architectures trained on a noisy dataset are studied. In addition, the external noise of the gathered dataset is reduced by exploiting an unsupervised method called Isolation Forest, and the new training results are examined. Based on the experiments, high recognition accuracies were achieved by training two states of the art networks on the noisy dataset, and the obtained results were slightly improved by using the proposed noise reduction framework. Finally, a demonstration application is provided to show the capability and the performance of the proposed approach.