Transfer Learning Inspired Fish Species Classification

Abstract

Machine learning techniques enable systems to learn Important representations from input Image data. Convolutional neural networks (CNNs) are a specific implementation of machine learning techniques and are able to create expressive representations from the input image. Hence CNNs are well suited for image processing operations such as classification, clustering, and object detection, etc. The creation of a new effectual deep CNN model involves an extensive training phase. This requires very large datasets, huge computation environments, and longer execution time. Several established deep CNNs are readily available. These networks are pre-trained on massive databases of images. VGG, ResNet, and InceptionResNetVZ are the leading pre-trained CNN models currently being used in numerous image-processing studies. Possibly we can transfer knowledge learned from such models in order to address challenges in different domains. This can be achieved by repurposing a deep CNN model as a feature generator to produce effective features for content based information retrieval applications. This research work proposes a technique for recognizing fish using deep convolutional neural networks such as ResNet-50, InceptionResNetVZ, and VGG16 that have been pre-trained using transfer learning.