Affiliations: School of Mathematics, Statistics and Computer Science, College of Science, University of Tehran, Tehran, Iran
Correspondence:
[*]
Corresponding author: Hedieh Sajedi, School of Mathematics, Statistics and Computer Science, College of Science, University of Tehran, Tehran, Iran. E-mail: [email protected].
Abstract: Concept detection for a collection of images is an important topic and has recently been an emerging area of studies. Facing an imbalanced dataset is a great challenge in concept detection that has not yet been adequately investigated. To cope with this challenge, this paper proposes an image concept detection system based on the Convolutional Neural Network (CNN) method. The proposed method consists of three stages. At the first stage, a new algorithm is proposed to enhance the batch sampling in the CNN. At the second stage, some augmentation methods are used to improve learning process in CNN and at the final stage, a new ensemble of balanced convolutional neural network is presented in order to detect the concepts of images. Using Caltech-101 image dataset, the experimental results demonstrate the effectiveness of the proposed framework for concept detection in imbalanced datasets.
Keywords: Concept detection, deep learning, convolutional neural networks, imbalanced dataset
