Paper Title
Transfer Learning or Custom CNN Network for Special Cases of Image Classification

Abstract
Convolutional Neural Networks (CNNs) have outperformed the traditional fully connected ANN (Artificial Neural Network) in the field of image analytics including image classification, localization etc. In this paper we have investigated if always a transfer learning based very deep CNN such as Resnet 50 or MobileNetV2 are the best choices for all kind of image classification tasks or there can be a room of optimization and a hand-crafted network works better. It has been found that while a 50 CNN layer network such as Resnet50 gave only 34% accuracy, a custom 11 CNN layered networkgave 59% accuracy and most importantly another custom network with 6 CNN layers provided an accuracy of 66%, whereas, the MobileNetV2network well known for its capability of processing low resolution images gave an accuracy of only38%. This clearly shows that not always a very deep neural network trained on a huge dataset like ImageNet performs best. Moreover, we have also investigated the effect of batch size on accuracy for a few of these networks.