Has computer vision become a sub-field of deep learning?

Computer vision is not strictly a sub-field of deep learning, but deep learning has become the dominant approach for solving many computer vision problems. Computer vision encompasses a broad range of techniques for enabling machines to interpret visual data, including traditional algorithms for tasks like edge detection, feature matching, and image segmentation. However, since the rise of deep learning—particularly convolutional neural networks (CNNs)—the field has increasingly relied on neural networks to achieve state-of-the-art results in tasks such as image classification, object detection, and semantic segmentation. While deep learning is now central to many applications, computer vision remains a distinct discipline with its own principles and methods that extend beyond neural networks.

The shift toward deep learning in computer vision began around 2012 with the success of AlexNet in the ImageNet competition, which demonstrated CNNs’ ability to outperform traditional methods. For example, tasks like object detection, which once relied on handcrafted features (e.g., Haar cascades or HOG descriptors) and classical machine learning models (e.g., SVMs), are now commonly addressed using architectures like YOLO, Faster R-CNN, or RetinaNet. Similarly, image segmentation has moved from graph-based algorithms (e.g., GrabCut) to deep learning models like U-Net or Mask R-CNN. These neural networks automate feature extraction, reducing the need for manual engineering and improving accuracy on complex datasets. However, traditional techniques still play roles in scenarios where data is scarce, computational resources are limited, or interpretability is critical—such as medical imaging pipelines that combine edge detection with CNNs.

While deep learning dominates research and industry applications today, computer vision is not subsumed by it. For instance, 3D reconstruction often uses structure-from-motion or SLAM algorithms that rely on geometric principles rather than neural networks. Similarly, real-time augmented reality systems might combine classic camera calibration techniques with deep learning for object tracking. Developers also frequently blend approaches: OpenCV, a staple computer vision library, is still widely used for preprocessing (e.g., noise reduction, perspective correction) before feeding data into a neural network. The field’s diversity ensures that while deep learning is a core tool, computer vision remains a multidisciplinary area integrating optics, signal processing, and traditional algorithms alongside modern neural networks.