Fish Classification Data Science Project

Overview

This repository contains code and documentation for the Fish Classification project, focusing on fish classification using a deep learning approach. The project involves training models to classify fish images into 9 different classes and different experiments.

Files

All files contain proper documantation, overview:

Classes

• CNN_Fish - model for fish classification
• FishDataset - dataset class to retrieve images

Model Train and Testing

• train - train the model using K-Fold cross validation
• utils - functionf for reading and train/test splitting of data

Pretrained Models

• model_architectures_main - main for using pretrained models as classifiers of images
• model_architectures_train_test - train,test,validate pretrained models
• model_architectures_utils.py - getters for different models and classifiers

Classifiers with Feature Extractor models

• feature_extractor_main - main for using pretrained models as feature extractor

Dataset

Link to Kaggle Dataset

Structure

The dataset comprises 9 labels of fish:

Original Images:

50 images for each fish (except one class with 30 images) RGB images with varying sizes

Training and Evaluation

Initial Model

We initially used original images with k-fold Cross Validation (k=5). For each fold: 80/20 train-validation/test split. The best-performing fold achieved the following results:

• Train Accuracy: 95.3%
• Validation Accuracy: 66.7%
• Test Accuracy: 54.7%

Examples of correctly/incorrectly/uncertain classified images:

Improvements

To enhance model performance, we implemented the following changes:

• Used augmented images (7 new roated images for each image)
• Hyperparameter tuning (learning rate, batch size, epochs)
• Data preprocessing (image normalization) Results for the best fold after improvements:

• Train Accuracy: 97.9%
• Validation Accuracy: 83.8%
• Test Accuracy: 84.9%

Examples of correctly/incorrectly/uncertain classified images:

Iterative Test-time Augmentation (ITA)

We experimented with ITA, achieving improved results but not surpassing the model without ITA:

• Train Accuracy: 99.9%
• Validation Accuracy: 81.2%
• Test Accuracy: 80.2%

Examples of correctly/incorrectly classified images:

Class Addition

A new class, "Janitor Fish," was added:

Results were slightly worse but still good performance:

• Train Accuracy: 93.5%
• Validation Accuracy: 80.5%
• Test Accuracy: 79.2%

Model Comparison

We compared the performance of different deep learning models:

VGG19:

• Validation Accuracy: 26.09%
• Test Accuracy: 26.74%

ResNet18:

• Validation Accuracy: 98.55%
• Test Accuracy: 100%

DenseNet121:

• Validation Accuracy: 98.55%
• Test Accuracy: 100%

ResNet50:

• Validation Accuracy: 98.55%
• Test Accuracy: 100%

Classical ML Algorithms:

We used ResNet18 as a feature extractor model and applied Random Forest, SVM, and KNN classifiers:

Random Forest:

• Validation Accuracy: 97.1%
• Test Accuracy: 100%

SVM:

• Validation Accuracy: 98.55%
• Test Accuracy: 100%

KNN:

• Validation Accuracy: 98.55%
• Test Accuracy: 100%

Fast Learning Experiment

We conducted experiments to assess how quickly models learn the dataset. ResNet18 demonstrated the best performance in terms of fast learning.