Implementation of low sampling rate algorithm

Abstract

Compressed learning integrates signal processing together with machine learning for inference from a signal with minimal measurements from a linear projection. Traditional image classification models learn directly from training images that are only processed for image augmentation and reshaping to fit the models. When these models are trained with a large number of images, it is highly inefficient as it could take extremely long durations of time to train or require a very powerful computer to complete the process. From previous reports that I reviewed for my literature review in chapter 2, some methods have been suggested to tackle this issue. The suggested solutions includes the usage of compressed learning and compressed sensing, which generally follows the same idea which is to reduce the amount of samples required to be taken from a signal or image to achieve an objective. In chapter 3, the theory of compressed sensing and compressed learning have been described in detail to show how they are made use of to achieve low sampling rates. In chapter 4, I briefly explained the theory of machine learning and proposed the most accurate image classification model within my means, Inception ResNetV2, that was able to achieve about 99% classification accuracy when trained on non-compressed domains. Furthermore, I described in detail the theory behind Principal Components Analysis (PCA), which I made use to compress the images for training and testing purposes. In chapter 5, I first reviewed the results of the machine learning model by testing it against images in the uncompressed domain, ensuring that the model is still competitive by achieving a precision of 97%. After this, I tested the image classification model against a test set with images in the compressed domain and achieved a high precision of 95%. Lastly, I reviewed the images that were wrongly predicted and provided suggestions for future works that could potentially improve the model further.