Update README.md

README.md

@@ -1,14 +1,18 @@
 
-# Image Data Augmentation: Low Rank Approximation Using the Singular Value Decomposition
+# On Image Data Augmentation: Training CNNs with Low-Rank Approximation of Images using the Singular Value Decomposition
 ---
 ##### Kennette Basco 
-##### Master's project 
+##### Master's Thesis 
 ##### New York Institute of Technology 
 ##### Department of Engineering and Computer Science 
 
 ---
-Data augmentation, particularly in image processing/computer vision, is a useful regularization method in training deep convolutional neural networks (CNN). It entails introducing variations in the training data that aid the CNN model to recognize patterns and spatial properties more effectively. Common techniques include: random alterations in orientation, rotation, scale, addition of pixel value noise, etc. Another potential addition to image augmentation techniques is using a low-rank approximation of images through singular value decomposition (SVD) to compress images while retaining crucial pixel information. 
-Regularization is any method that decreases the optimization performance (some metric on the training data) that improves generalization performance (some metric on the testing data). Since the goal of machine learning is to predict data that has not been seen yet, regularization of models is crucial to avoid overfitting the training data. The regularization methods on images aforementioned are only active during training, i.e., at inference time, the images are not distorted.
-The regularization effect of SVD will be determined in the binary and multiclass classification settings using a trained deep learning model.
+##### ABSTRACT
+This study investigates enhancing the generalization of Convolutional Neural Networks (CNNs) for low-resolution images using Singular Value Decomposition (SVD) augmented images. It provides a theoretical overview of SVD and a method to control compression through hyperparameters: energy factor and skip threshold. Methods involve preprocessing, training models, and benchmarking validation and testing accuracy on various datasets (grayscale and RGB). Experiments using SVD augmentation show degraded validation and testing accuracy. In summary, while SVD augmented data shows theoretical promise, it does not alleviate the training-production data skew and even increases the signal-noise ratio. Rigorous hyperparameter tuning of the energy factors and skip threshold may regularize CNNs, but the computational resources required outweigh the marginal gains in regularization, if at all. This study highlights the limitations of SVD based augmentation and underscores the need for alternative regularization techniques in deep learning.
+Keywords: convolutional neural networks, singular value decomposition, regularization, adversarial learning.
 
+---
+##### SVDNet
+SVDNet is a feature extracted and fine-tuned pre-trained VGG16 model using SVD-augmented images
 
+![MNIST](/resources/mnist.gif)![PATCH](/resources/patch_camelyon.gif)![MALARIA](/resources/malaria.gif)![CATSDOGS](/resources/cats_vs_dogs.gif)