Zero-Shot Learning is a machine learning method, where the model tries to generalize or classify concepts or objects which are unseen during the model training. Machine vision tasks have significantly advanced with Zero-Shot Learning and Generalized Zero-Shot Learning, enabling models to classify images without prior training data for specific classes. Zero-Shot Learning allows models to identify unseen classes by leveraging auxiliary semantic information such as textual descriptions and pre-trained models; for instance, Contrastive Language-Image Pre-training, Vision Transformers etc. This is achieved by aligning visual features with semantic embeddings in a shared space, mimicking human-like recognition. Generalized Zero-Shot Learning addresses some of the issues face by Zero-Shot Learning, by allowing recognition of both familiar and novel categories simultaneously, aiming for more real-world applicability. Despite Generalized Zero-Shot Learning’s progress, issues like bias amplification towards seen classes, catastrophic forgetting, domain shift, and mode/posterior collapse persist. Current research explores generative models including Generative Adversarial Networks and Variational Auto-Encoders, meta-learning, and incremental learning to mitigate these challenges, striving for robust and adaptable machine vision systems in complex, unstructured environments. This research paper objectively analyzes the current challenges and methods in Zero-Shot Learning tasks. The analysis start by formally defining the Zero-shot and Generalized Zero-Shot Learning problems, and do a comprehensive investigation on the recent models that addresses aforementioned challenges. This study and analysis focuses on identifying the research gaps of Zero-Shot learning in computer vision tasks. This paper further provides open questions for future research.