Machine Learning Solutions for Classification and Regions of Interest Analysis on Imbalanced Datasets

This dissertation investigates machine learning algorithms: Linear Discriminant Analysis (LDA) and Generative Adversarial Networks (GANs) to address real-world tasks suffering from imbalanced problems efficiently and robustly. The proposed LDA variants and GANs variants in this dissertation are used for classification and regions of interest analysis tasks with different severely imbalanced inputs.

This dissertation firstly contributes to addressing imbalance problems with LDA-related methods for binary-label and multi-label classification tasks. The traditional LDA has been widely used as a pre-processing step to enhance the efficiency and performance of sequential classifiers on datasets satisfying the Gaussian distribution. To extend the traditional LDA for uneven and diverse inputs, We initially introduce weight factors into the definition of scatter matrices based on a novel probabilistic saliency estimation method as saliency-based weighted LDA. Usually, the weight factors are exploited based on specific metrics with the label or feature correlation of input data. The redefinition can balance the sample contribution to mitigate the influence of outlier samples for binary-label classification tasks. The experimental performance of the proposed methods has been assessed over six publicly facial datasets, which demonstrates a robust improvement compared to the competing methods.

To address more complicated imbalanced problems widely existing in multi-label classification tasks, this dissertation introduces a saliency-based multilabel LDA framework. The proposed framework extends the probabilistic saliency estimation method for multi-label classification tasks based on multilabel LDA scatter matrices to alleviate the performance degradation caused by imbalanced problems. Six kinds of weight factors are obtained by the probabilistic saliency estimation method and the exploration of prior information with six metrics. The experimental results of the proposed methods over 17 imbalanced datasets show remarkable performance improvements compared to competing methods with seven quantitative evaluation metrics

Alternative approaches to LDA include various deep learning models which have been explored for various regions of interest analysis tasks which may include imbalanced datasets in many applications. Especially, the GANs-related methods with appropriate deep neural networks can be used to address various regions of interest analysis tasks suffering from imbalance problems in computer vision areas efficiently and effectively, due to its excellent functionality of restoring balance in the problem. Another contribution of this dissertation is to explore GANs-based methods for two regions of interest analysis tasks with severely imbalanced inputs.

This dissertation investigates optimized Deep Convolutional Generative Adversarial Networks (DCGANs) to edit specific facial attributes, such as occlusions. The proposed method utilized a pre-trained DCGANs and an optimization loss function to detect the occluded facial regions and in-paint with corresponding facial attributes. The pre-trained DCGANs is trained with occlusion-free facial images to distinguish facial attributes and occlusions during the inference stage with the optimization function. The visual experimental results have shown that the proposed method can detect the required facial occlusions and then successfully in-paint them with the corresponding facial attributes.

Besides, a conditional Generative Adversarial Network (cGANs) based framework is introduced to detect anomalous regions (e.g., cracks) on pavement images efficiently and effectively in this dissertation. Such a task is also considered a binary semantic segmentation problem with imbalanced data due to uneven distribution between the number of the required anomalous region pixels and the background pixels. The proposed cGANs-based method consists of a UNet-based generator part for a multiscale feature representation, a discriminator part for real pairs and fake pairs judgment, and a novel auxiliary network for a refined feature representation. To increase performance while avoiding increasing network and computational complexity, the proposed framework is trained alternatively in two stages. The proposed methods have shown the effectiveness of GANs-based methods and their robustness in tackling binary semantics segmentation with severely imbalanced inputs through extensive experiments over six benchmark datasets.