Corrosion and cracks in industrial pipelines cause safety hazards, economic losses, and environmental damage. Timely detection is crucial for avoiding catastrophic failures. Traditional nondestructive testing (NDT) methods require human interpretation and often lack reliability in harsh environments. The purpose of this review is to analyze deep learning techniques used for pipeline defect detection and evaluation. The study focuses on convolutional neural networks (CNN), vision transformers (ViT), and image segmentation models applied to corrosion and crack identification. Different data sources are considered, including visual inspection, ultrasound imaging, radiography, and drone-based monitoring. The review compares available datasets, their limitations, and labeling difficulties. Inspection scenarios such as real-time monitoring, underwater pipelines, and high-temperature environments are highlighted. Challenges related to noise, occlusion, illumination changes, and generalization are discussed. The analysis also covers decision-making systems that support risk assessment and maintenance planning. The findings show that deep learning significantly improves defect detection accuracy compared to traditional approaches. However, model performance depends on data quality and domain adaptation. The review concludes that integrating multimodal sensing, real-time inference, and explainable artificial intelligence (XAI) is essential for practical deployment. Future research must focus on robust dataset development, uncertainty estimation, and autonomous inspection systems.