NEURAL NETWORK FISH CLASSIFIER

Introduction

This is an implementation of a convolutional neural network that is able to classify groups of fish. The dataset has got four different fish classes, each one composed of 1000 images, and they have been obtained from a Kaggle dataset (see https://www.kaggle.com/crowww/a-large-scale-fish-dataset).

Best Results

• Accuracy and Loss Evolution

• Confusion Matrix

Hyperparameters and Parameters

• Dataset initializer

Name	Definition	Value
Training percentage	Amount of files for training in %	80%
Validation percentage	Amount of files for validation in %	20%

• Neural network Compilation

Name	Definition	Value
Loss function	Loss function used during training	Categorical Cross Entropy
Optimizer	Optimizer used during training	Ada Delta

• Training and EarlyStopping

Name	Definition	Value
Number of Epochs	Maximum number of iterations	30
Steps per Epoch	Total steps of each epoch executed	100
Batch size	Number of files loaded during training steps	30
Patience	Maximum tries to improve val_accuracy before exit	3

• Data Augmentation

Name	Definition	Value
Target size	Width and height for each image loaded	250w, 250h
Rescale	Rescale used during image loading process	1./255
Rotation range	Rotation range for data augmentation	30
Zoom range	Zoom range for data augmentation	0.7
Width shift range	Width swift range for data augmentation	0.1
Height shift range	Width swift range for data augmentation	0.1
Brightness range	Brightness range for data augmentation	(0.2, 0.8)
Horizontal flip	Horizontal flip for data augmentation	True
Vertical flip	Vertical flip for data augmentation	True

• Models

Name	Structure
ModelDefinitionOneConv	Conv2D(filters=32, kernel=(2, 2), activation=relu) MaxPooling((2,2)) Dropout(0.25) Flatten() Dense(64, activation=relu) Dropout(0.5) Dense(32, activation=relu) Dense(4, activation=softmax)
ModelDefinitionTwoConv	Conv2D(filters=64, kernel=(2, 2), activation=relu) MaxPooling((2,2)) Dropout(0.25) Conv2D(filters=128, kernel=(2, 2), activation=relu) MaxPooling((2, 2)) Dropout(0.25) Flatten() Dense(128, activation=relu) Dropout(0.25) Dense(64, activation=relu) Dense(4, activation=softmax)
ModelDefinitionThreeConv	Conv2D(filters=32, kernel=(2, 2), activation=relu) MaxPooling((2,2)) Dropout(0.25) Conv2D(filters=64, kernel=(2, 2), activation=relu) MaxPooling((2, 2)) Dropout(0.25) Conv2D(filters=128, kernel=(2, 2), activation=relu) MaxPooling((2, 2)) Dropout(0.25) Flatten() Dense(64, activation=relu) Dropout(0.5) Dense(32, activation=relu) Dense(4, activation=softmax)

Name	Loss	Accuracy	Validation Loss	Validation Accuracy
ModelDefinitionOneConv	1.3097	0.4203	1.1064	0.6775
ModelDefinitionTwoConv	1.2537	0.5282	1.0885	0.6775
ModelDefinitionThreeConv	1.3853	0.2657	1.3765	0.3663

Authors

• losedavidpb: https://github.com/losedavidpb
• SergioULPGC: https://github.com/SergioULPGC