An efficient multi-neural network ensemble model for image classification

Abstract

Image classification is an artificial intelligence (especially deep learning) technique used to classify an image into specific categories or classes. Today, it is one of the cornerstones of computer vision and is of vital importance in many fields. For example, categorizing with high accuracy medical images into disease classes enables more efficient and accurate diagnosis. To achieve high accuracy in image classification tasks has encouraged the development of methods such as CNN. In addition, some methods such as ensemble technique and Transfer Learning etc. are commonly used for this objective. However, while trying to achieve high accuracy, other important parameters such as training time must also be considered. Therefore, especially Transfer Learning method is widely applied in image classification to reduce training time and enhance model efficiency. Even though transfer learning with pre-trained models such as AlexNet, VGG16, and DenseNet121 is widely used, when using these models for some image dataset, it demands a great amount of training time to reach high accuracy. The objective in this thesis is not only to increase accuracy but also to reduce training time for image classication tasks. Hence, it is proposed a model for image classification that incorporates five deep learning architectures with an ensemble technique. The proposed model consists of one MLP-based network and four CNN-based networks where one of them is a network that we call the auxiliary network. The auxiliary network is designed to recognize misclassified images in order to increase the accuracy of the model. The proposed model is tested on an image dataset called CIFAR-10. Then, it is compared the performance of the proposed model with pre-trained structures such as AlexNet, VGG16, and DenseNet121 on taking into account training time, the number of parameters, and accuracy. The results show that the proposed model outperforms pre-trained models in terms of achieving high accuracies and requiring less training time on CIFAR-10 dataset. The proposed model requires 15,38%, %10, and %87.78 of the training time of Alexnet, VGG16 and DenseNet121 to achieve %80 accuracy., respectively. While the proposed model achieves 85% and 90% accuracy, AlexNet and VGG16 cannot. In addition, it achieves 90% accuracy in 38.23 min, whereas DenseNet121 – more efficient than the other two pre-trained models - only reaches 87% accuracy in over three hours.