Classifying scenes is no difficult task for humans. It is a learned process. Inspired by biological neurons in the visual cortex, the concepts of convolutional neural networks (CNNs) have evolved since the 1980s. With the increase in computational power and data availability today, deep neural networks can be trained to classify images. Training a classifier that accurately classifies images based heavily depends on the parameters used in the CNN model, the number of convolutional layers used, the number of iterations(backpropagations/epochs), number of filters used, kernel size, the accuracy of the labeled data, and the size of the dataset. In this paper, a dataset of 17034 images is used to train a CNN image classification model. In addition to the CNN model, transfer learning models are trained on pre-trained networks like ResNet, VGG16, and IceptionV3. All models are experimented on using TensorFlow and Keras libraries, and the classifier that predicts the class of the image scene with 91.5% accuracy, built on top of the ResNet50 architecture trained without rescaling image size, is recommended. The different parameters and architectures that affect the accuracy are the computational time of the outcome are also discussed.
Please sign in or create an account to give a rating or comment.