Authors: Bagal Prem Sandip, Prof. S. K. Said, Prof. M. C. Kolse
Abstract: Computer vision enables machines to extract and interpret meaningful information from images and video. It has evolved from hand-crafted image processing and feature descriptors to deep learning systems that now perform classification, detection, segmentation and recognition at or near human level on many benchmarks. This paper reviews the field in a structured way. It begins with classical foundations including edge detection, scale-invariant keypoints and histogram-of-gradient features, then traces the deep learning revolution enabled by large labelled datasets and convolutional neural networks, covering residual networks, region-based and single-stage object detectors, fully convolutional and instance segmentation, and vision transformers. Recent foundation models that align images with language or segment arbitrary objects from prompts are also discussed. Applications in autonomous driving, medical imaging, robotics, surveillance and augmented reality are summarized. A generic five-layer computer vision pipeline from image acquisition to decision output is proposed, along with research gaps, objectives, a comparison of architectures and an analysis of challenges such as data dependence, robustness, bias, privacy and computational cost. The review concludes that computer vision is moving toward general-purpose, multimodal models that must also become efficient, robust and trustworthy.
DOI: https://doi.org/10.5281/zenodo.24220646
