Deep learning approaches in computer vision: a comprehensive survey

Deep learning now underpins modern computer vision across various applications, including classification, detection, segmentation, retrieval, and generation. This survey first revisits foundational architectures—CNNs, early RNN-based vision pipelines, GANs, and the first Vision Transformers—then traces the shift to today’s practice (2023–2025): vision-language models (e.g., CLIP), large-scale self-supervised pretraining (e.g., DINOv2), promptable foundation segmentation (SAM), and diffusion-based generators (DDPM, latent diffusion). We distill the core design principles and training patterns behind these systems, analyze their strengths and limitations relative to classical approaches, and identify deployment-relevant concerns: compute efficiency, data governance, safety, and evaluation rigor. Looking ahead, we highlight parameter-efficient adaptation of vision foundation models (prompting/adapters), reliable few-shot protocols, and hardware-aware neural architecture search that meets latency and energy budgets. These priorities define concrete knowledge gaps that delineate the current research frontier. In this 2025 update, we introduce a concise survey of quantum–classical hybrids for vision, covering QCNNs, QViTs, and QGANs, and highlight their promise in small-data or tight-parameter regimes.