Streamlining Choice of CNNs and Structure Framing of Convolution Layer

Convolutional neural networks (CNNs) are a kind of deep neural networks which were designed from the biologically driven models. Researchers focused on how humans perceive an image in the brain. As an image is passed through different layers in the human brain, in the same way, CNNs have many layers. In this paper, the structure of CNN is described, guidelines on the design of the convolution layer and decision making on when to use a pre-trained CNN model with transfer learning and when to design our custom architecture CNN model. This will help future researchers in a quick start with CNN modeling. Experimentation is done on two popular image datasets, i.e., CIFAR-100 and Stanford clothing attribute dataset, where CIFAR-100 is a clean dataset of 60,000 images belonging to 100 classes and Stanford clothing attribute dataset is highly noisy and imbalanced data as it has uneven distribution of samples for different attributes and many of the samples do not have a clear distinction between the classes resulting in overlapping training data. Four CNNs were designed, where two models were pre-trained CNN models and two were customized CNN models and compared their performance on image classification task and treatment of the missing data in the dataset. Based on this comparison and related study, we framed the guidelines for designing a convolution layer and making choice between using a pre-trained (transfer learning) or customized CNNs.