<p>The increasing need for automated quality control techniques has been driven by the rapid advancement of smart manufacturing systems within the Industry 4.0 framework. Machine vision using deep learning technologies has emerged as a key enabler of streamlined inspection and decision-making workflows across various production environments. This systematic review analyzes the evolution of deep learning models from traditional Convolutional Neural Networks (CNNs) to transformer-based architectures, examining how these developments address challenges in defect detection and classification on modern industrial production lines. Several key findings emerge from our analysis. First, while CNNs continue to dominate due to well-established performance in local feature extraction, transformer models demonstrate superior accuracy for complex defect geometries through global contextual reasoning, albeit at higher computational costs. Second, real-time detection methods such as YOLO have proven viable for high-velocity production environments, though they typically sacrifice accuracy for speed. Third, integration with edge computing, digital twins, and IoT infrastructure is essential for developing scalable quality control systems. Additionally, ongoing challenges include data format inconsistency, model interpretability, domain transfer limitations, and the gap between controlled development conditions and actual factory environments. This review concludes by outlining future research priorities: developing hybrid CNN-Transformer models, establishing standard evaluation criteria, creating resource-efficient edge applications, and building collaborative human-AI frameworks, thereby providing guidance for researchers and engineers advancing automated visual inspection systems.</p>