Data augmentation plays a crucial role in improving deep learning models, yet conventional approaches often result in feature loss or biased learning. This thesis introduces novel enhancement techniques that address these challenges in multiple domains. First, Random Slices Mixing Data Augmentation (RSMDA) enhances feature diversity by strategically combining image slices while leveraging label smoothing, improving model generalisation. Second, RandSaliencyAug (RSA) balances feature loss and contextual information retention by selectively occluding salient regions using six new strategies, outperforming existing occlusion-based methods. Third, KeepOriginalAugment integrates salient regions into non-salient areas, striking a balance between data diversity and information preservation through optimised placement strategies. We evaluated these methods in FashionMNIST, STL10, CIFAR10, CIFAR100, TinyImageNet, ImageNet, and VOC 2007. In addition, we provide depth analysis and comparison including class activation maps (explainability), robustness, and time complexity. Furthermore, FaceSaliencyAug and FaceKeepOriginalAugment mitigate geographical, gender, and stereotypical biases in computer vision models, improving fairness and diversity. We also explore and propose two novel data augmentation techniques from an image-mixing perspective: Noise Addition (NA) and Partial Mix (PM). These techniques were evaluated on FFHQ, WIKI, IMDB, LFW, and UTK Faces datasets, as well as diverse datasets, demonstrating significant improvements in diversity. We evaluate their impact on mitigating gender bias across CEO, Engineer, Nurse, and School Teacher datasets, using the Image-Image Association Score (IIAS) in convolutional neural networks (CNNs) and vision transformers (ViTs). In addition, we introduce a Saliency-Based Diversity and Fairness Metric to quantify bias reduction across datasets. Our approaches are rigorously tested on CNNs and ViTs with extensive evaluations on facial recognition datasets, proving their effectiveness in enhancing both model performance and fairness.