Journal of Engineering Research and Reports

Object detection and image segmentation are foundational tasks in computer vision, enabling machines to localise, classify, and delineate objects within images and video streams. Over the past decade, deep learning has transformed these fields beyond recognition, delivering performance gains that consistently surpass earlier handcrafted-feature approaches. This article presents a comprehensive narrative review of deep learning-based methods for object detection and segmentation, tracing the evolution from seminal convolutional architectures to contemporary transformer-based frameworks and foundation models. The literature for this review was identified through searches conducted in Web of Science, Scopus, Google Scholar, and PubMed.  The review examines two-stage and one-stage detection paradigms, anchor-based and anchor-free detectors, semantic and instance segmentation, panoptic unification, self-supervised representation learning, data augmentation strategies, and transfer learning. Key benchmark datasets and evaluation metrics are discussed, as are applications across autonomous driving, medical image analysis, remote sensing, and small-object detection. The article concludes by identifying persistent challenges—including small-object detection, domain shift, annotation scarcity, and computational efficiency—and by outlining directions likely to define the next phase of progress in the field. The integration of temporal information for video-domain detection and segmentation continues to develop. And the fundamental challenge of robust generalisation across diverse, uncontrolled environments—which benchmarks have never fully captured—remains the field's most important open problem.