🌸 Comparative Analysis of Segmentation Performance

Overview

This project delves into the applications of deep learning Convolutional Neural Networks (CNNs): DeepLab v3+ and a CNN inspired by U-Net called Floral-Net, tasked with segmenting flowers and backgrounds from the Oxford Flower Dataset.

Methodology

This study addresses imbalanced data in the Oxford Flower Dataset by carefully preprocessing images and labels. Using MATLAB, missing labels are removed, resulting in 846 image-label pairs. Transfer learning with DeepLabv3+ using ResNet18 and ResNet50 is employed, alongside a unique Floral-Net architecture inspired by prior works. Floral-Net integrates incremental dropout, selective filter adjustments, and weighted pixel classification to mitigate class imbalance. This comprehensive approach aims to improve semantic segmentation of flower images.

Dataset

The Oxford Flower Dataset (17 Classes) was used.

Training Options

DeepLabv3+ Initialised with ResNet18/50 weights

Option	Value
Optimization Algorithm	sgdm
Learning Rate Schedule	piecewise
Learning Rate Drop Period	6
Learning Rate Drop Factor	0.1
Momentum	0.9
Initial Learning Rate	0.01
L2 Regularization	0.005
Validation Data	dsVal
Max Epochs	10
Mini-Batch Size	4
Shuffle	every-epoch
Checkpoint Path	tempdir
Verbose Frequency	10
Validation Patience	4

Floral-Net

Option	Value
Optimization Algorithm	Adam
Initial Learning Rate	0.001
Learning Rate Schedule	Piecewise
Learning Rate Drop Factor	0.1
Learning Rate Drop Period	10
L2 Regularization	0.0001
Maximum Epochs	10
Mini-Batch Size	16
Data Shuffling	Every Epoch
Validation Data	dsVal
Validation Frequency	10