Scene classification based on convolutional neural network

Abstract

Convolutional Neural Network(CNN) has been widely used in image recognition and classificaiton. The objectives of this project is to implement the mupltiple CNNs on MIT Indoor67 dataset, to evaluate their performance, and therefore gain first hand experience on transfer learning. First, Place205-VGG CNN model has been used in order to evaluate its performance. Later, after performance evaluation, a relatively new technique CAM has been implemented on CNN, as a result, a heat map will be generated so that human can visualize and indirectly understand the relative importance of feature information learned by CNN. This manoeuvre enables a deeper understanding of CNN’s learning aspect. Second, pre-trained ResNet-152 has been used for fine- tuning, by freezing some low-level layers and training the final classifier, a better classification accuracy is obtained.