A Comparative Analysis on Architectures of UNet Family for Image Segmentation

Abstract

Image segmentation is an important visual task that involves categorizing each pixel within an image, assigning it a specific label that defines its content. The advancement of deep learning has revolutionized this field, with Convolutional Neural Networks (CNNs) marking a significant improvement in accuracy and efficiency. These models learn from the raw image data to the segmentation labels in an end-to-end manner, enabling them to produce precise segmentation outcomes with minimal human intervention. Among the spectrum of segmentation models, the UNet architecture has emerged as a standout performer, renowned for its robust performance and adaptability. It has not only set a high benchmark but also inspired a plethora of enhancements and adaptations to cater to various applications. This paper presents a systematic analysis of UNet, hoping to share the UNet model with its extensions. So that beginners can have a preliminary understanding of this famous family of deep-learning network models. In this paper, the author first introduces the basic UNet model. Then introduce eight classic UNet variants. This paper will guide the reader to understand the compilation and improvement principles of the UNet family and provide ideas and a foundation for the reader's subsequent research on UNet.