ML-For-Fish-Recognition

Created for my Master's Thesis.

Name: Giannis Kostas Georgiou

Student ID: 20090232

Supervisor: Dr Peter McBurney

The Android App files and info are located at:

• https://github.com/jcgeo9/ML-For-Fish-Recognition-App

Repository Contents

This repository is divided into the following directories:

• Dataset Preparation, Model Training and Model Testing Python Code
• Saved Binary and Multi-Class Classification Models

Datasets

The datasets used in this repository and their corresponding kaggle repositories are:

• Fish-No Fish
  • https://www.kaggle.com/giannisgeorgiou/fish-or-no-fish-simple-images
• Species
  • https://www.kaggle.com/giannisgeorgiou/fish-species
• Fish-No Fish with Species Images
  • https://www.kaggle.com/giannisgeorgiou/fish-or-no-fish-species-images
• Fish-No Fish with Combined Images
  • https://www.kaggle.com/giannisgeorgiou/fish-or-no-fish-simple-and-species-images

Instructions on Python Files

Instructions on how to use the python files are provided inside each file as a comment section on top but in general one should:

1. Download the dataset of choice (Can be either one of the above or another binary or multi-class dataset)
2. Download the directory corresponding to the downloaded dataset
3. Use the "/dataset" directory to convert and store the dataset
4. Use the "/model_training" directory to train models and find the most suitable
5. Use the "/model_testing" directory to test trained models of his choice

If one wishes to avoid training models and wants to obtain the trained ones they are located in "/Saved-Models" directory and can be used after the following:

1. Download the model directory
2. Loading the model using the following command

#loads the saved model from the path specified

model=tf.keras.models.load_model(model_path)

If one chooses to use a model which is not stored here, training models with these files may not produce a good result. Model accuracy and loss depends on the domain and the dataset, thus changing the dataset but not the model architecture will have a different result