Image Classification: Feature Selection, Data Augmentation, and Transferred Learning

College: New York Institute of Technology
Course: DTSC 870 - M01/Spring 2022
Advisor: Dr. Cao
Team: Micheal Trzaskoma and Hui (Henry) Chen

Introduction

Computer vision is a growing area of interest utilizing computers to evaluate visual input for various tasks, one of these tasks being image classification. When working with visual data for classification, one of the limitations is the lack of available labeled image data to train models. This project explores several methods to combat the limited image data which are feature modification and selection, data augmentation, and transfer learning. These methods were applied to convolution neural networks and support vector machines where applicable within experiments trying to optimize the accuracy of image classification. The two domains of classification explored in this project are medical data utilizing a MRI brain tumor dataset , and emotion recognition using the facial emotion recognition 2013 dataset.

Objective

To explore the techniques of feature selection, data augmentation, and transfer learning seeking to improve Machine/Deep Learning model (SVM and CNN) image classification accuracy on medical and emotion recognition.

Tools and Technologies

Fig [1]: Tools and Technologies Utilized Throughout the Project

Datasets

All datasets are available on Kaggle:

MRI Brain Tumor: here
FER-2013: here

Fig [2]: MRI Brain Tumor Image Class Distribution

Fig [3]: FER-2013 Image Class Distribution

Experiment Results

The following visuals are our experiment results for SVM and CNN models.

MRI Brain Tumor Dataset

SVM

Fig [4]: MRI - SVM Performance On Raw Pixel Features

Raw Pixel Features (Without PCA)	PCA Variance	# of Features
	100%	279
	99%	245
	97%	206
196608	95%	179
	90%	131
	80%	77
	70%	46

Table [1]: Raw Pixel Data and PCA Features for MRI Dataset

Fig [5]: MRI - SVM PCA Experiment Accuracy Progress

Raw Pixel Features (Without PCA)	PCA Variance	# of Features (Without PCA)	LBP (P : 8, R : 1)	LBP (P : 16, R : 2)	LBP (P : 24, R : 3)
	100%	279	279	279	279
	99%	245	272	273	273
	97%	206	262	264	265
196608	95%	179	253	255	257
	90%	131	232	235	239
	80%	77	193	200	205
	70%	46	158	166	173

Table [2]: Raw Pixel Data and LBP Handcrafted Features and PCA

Fig [6]: SVM with PCA for LBP (P: 8, R: 1)

Fig [7]: SVM with PCA for LBP (P: 16, R: 2)

Fig [8]: SVM with PCA for LBP (P: 24, R: 3)

Fig [9]: MRI - SVM Experiment Accuracy Progress

CNN

Fig [10]: MRI - CNN Structure

Fig [11]: MRI - CNN Structure

Fig [12]: MRI - DenseNet169 Fine Tuning

Fig [13]: MRI - DenseNet169 Pre and Post Augmentation Confusion Matrix

Fig [14]: MRI - CNN Experiment Accuracy Progress

FER-2013 Facial Dataset

SVM

Fig [15]: FER-2013 - SVM Performance On Raw Pixel Features

Raw Pixel Features (Without PCA)	PCA Variance	# of Features
	100%	2304
	99%	904
	97%	425
2304	95%	256
	90%	104
	80%	32
	70%	13

Table [3]: Raw Pixel Data and PCA Features for FER-2013 Dataset

Fig [16]: FER-2013 - SVM PCA Experiment Accuracy Progress

Raw Pixel Features (Without PCA)	PCA Variance	LBP (P : 8, R : 1)	LBP (P : 16, R : 2)	LBP (P : 24, R : 3)
	100%	10	18	26
	99%	8	13	18
2304	97%	7	10	13
	95%	4	6	7
	86%	2	2	2

Table [4]: FER-2013 - Raw Pixel Data and LBP Handcrafted Features and PCA

Fig [17]: SVM with PCA for LBP (P: 8, R: 1)

Fig [18]: SVM with PCA for LBP (P: 16, R: 2)

Fig [19]: SVM with PCA for LBP (P: 24, R: 3)

Fig [20]: FER-2013 - SVM With HoG Accuracy

Fig [21]: FER-2013 - The Effective of K in SIFT

Fig [22]: FER-2013 - SVM With SIFT Feature Accuracy

Fig [23]: FER-2013 - SVM Experiment Accuracy Progress

CNN

Fig [24]: FER - CNN Structure

Fig [25]: FER-2013 - ResNet50 With Different Trainable Layers

Fig [26]: FER-2013 - VGG16 With Different Trainable Layers

Fig [27]: MRI - VGG16 Pre Augmentation

Fig [28]: MRI - VGG16 Post Augmentation

Fig [29]: FER-2013 - VGG16 Per Class Performance Utilizing 48*48 Pixels vs 224*224 Pixels

No Augmentation	Reflection and Rotation	Reflection and Translation
64.11%	62.52%	64.43%

Table [6]: FER-2013 - CNN Accuracy Per Augmentation Methods

Fig [30]: MRI - CNN Experiment Accuracy Progress

Reports

Final Report: here
Presentation Slide: here
Presentation Video: Coming Soon!
Bi-Weekly Reports: here

Team

Micheal Trzaskoma

Data Preprocessing
Data Augmentation
CNN Experiments

Hui (Henry) Chen

Data Preprocessing
Data Augmentation
SVM Experiments