Computer Vision

In recent years, one of the most transformative subfields of machine learning has been computer vision. With substantial breakthroughs in the early 2010s, this field has seen improvements in tasks ranging from image classification, retrieval (similarity search), and denoising to more recent generative aspects such as image generation and style transfer. These advancements have been made possible by deep learning-based approaches, which have revolutionized the field. This chapter focuses on deep learning-based approaches that enabled the aforementioned advancements in computer vision. We begin by providing an overview of deep learning from images, including convolutional neural networks (CNNs) and their applications in image classification, object detection, and semantic segmentation. We then turn our attention to more recent endeavours to perform computer vision tasks with non-convolutional neural networks, such as attention-based neural networks. We also discuss recent advancements in neural network architectures, including residual networks and densely connected networks, which have significantly improved the performance of deep learning-based approaches in computer vision. We examine the challenges of training deep neural networks, including overfitting and vanishing gradients, and discuss techniques such as regularization and batch normalization that have been developed to address these challenges. Furthermore, we highlight a collection of use cases that pinpoint both strong points and some drawbacks of the current technology. Deep learning-based approaches have transformed the field of computer vision, enabling significant improvements in a wide range of tasks. By providing an overview of deep learning from images, recent advances in neural network architectures, and a collection of use cases, this chapter serves as a sensible resource for researchers and practitioners in the field.