Comparitive Analysis Of Different Architecture of Convolutional Neural Network

Abstract

Deep Neural Networks (DNN) have widely been experienced on images and videos comprising of large dataset in order to make differentiation among the images and videos based on DNN’s artificial sense of recognition. The Convolutional Neural Network (CNN) architecture is largely based on multiple different types of convolutional layers containing multiple neurons and each neuron in a layer is designed towards feed-forward direction. CNN is able to learn features of unstructured data such as images, voice and videos during the training process of a model. This study compares the CNN architectures of VGG-16, ResNet50, Xception and DenseNet121. The findings of this paper shows that CNN performs well while classifying digital images. If the number of weighted layers increase in a CNN model it results in the higher level of object detection accuracy. On the other hand if more weighted parameters are included in the CNN then it requires more time for CPU and GPU in the training session. The Xception model is better than VGG-16, ResNet50 and DenseNet121 because it passes same input to the depth wise isolated blocks and later merges the output of these blocks as input for the final classification layer. The sizes of the model are directly proportional to the number of parameters involved which ultimately affects the performance of the models during the object recognition process. These all models are pre-trained. They require transfer learning for fine tuning in a new dataset. There is a limitation to the models that dataset must be in the RGB- system.