Towards Designing the Best Model for Classification of Fish Species Using Deep Neural Networks

EasyChair Preprint 7442

Abstract

Convolutional Neural Networks have become a powerful tool for classification since 2012. The winners of the ImageNet Challenge have been neural networks for a long time now. Computer vision applications mostly resort to neural networks. This paper extends its application to classify fishes of 23 different species using VGGNet algorithm. The fish images used for training the network is obtained from a live video dataset. We implemented the traditional VGG16 and proposed our scaled down version of it, the VGG8. The results of training these two algorithms were compared on the basis of micro average, macro average and weighted average. VGG16 surpassed VGG8 in almost all parameters but not by a large margin as proved in the results section. Smaller networks consume less memory and take lesser time for training as well as prediction. Hence, smaller networks can be used in simpler applications and can run on a less dedicated hardware setup having restrictions of memory and/or processing power. Thus, it forms a good comparison problem which has been addressed in the paper below.

Keyphrases: Convolutional Layer, Convolutional Neural Networks, Convolutional neural network architecture, Fish Detection, Fully Connected Layer, Pooling layer, Sonar image, Training Loss, VGGNet, activation function, computer vision, deep learning, fish specie, image classification, macro average, neural network, output layer, scaled down version, smaller network

@booklet{EasyChair:7442,
  author    = {Pranav Thorat and Raajas Tongaonkar and Vandana Jagtap},
  title     = {Towards Designing the Best Model for Classification of Fish Species Using Deep Neural Networks},
  howpublished = {EasyChair Preprint 7442},
  year      = {EasyChair, 2022}}