# TernausNet: U-Net with VGG11 Encoder Pre-Trained on ImageNet for Image Segmentation

**Authors:** Vladimir Iglovikov, Alexey Shvets  
**arXiv:** [1801.05746](https://arxiv.org/abs/1801.05746)  
**Date:** 2018-01-17  
**Summary page:** https://www.summarizepaper.com/en/paper/1801.05746v1/

## Key points

- Reliable image segmentation is crucial in computer vision, especially in medical imaging and self-driving vehicles.
- Automated treatment is necessary for accuracy and efficiency as manual segmentation is time-consuming and prone to errors.
- Fully Convolutional Networks (FCN) are one of the most successful deep learning methods for image segmentation.
- Pixel-wise image segmentation is a demanding task, especially for medical images and satellite images.
- Classical U-Net architectures composed of encoders and decoders are popular for such segmentations.
- Pre-trained neural networks on large datasets like ImageNet show better performance than those trained from scratch on small datasets.
- TernausNet: U-Net with VGG11 Encoder Pre-Trained on ImageNet for Image Segmentation improves the U-Net architecture by using the pre-trained encoder.
- The code and corresponding pre-trained weights are publicly available at https://github.com/ternaus/TernausNet.
- The authors compare three weight initialization schemes: LeCun uniform, the encoder with weights from VGG11, and full network trained on the Carvana dataset.
- This network architecture was part of the winning solution (1st out of 735) in the Kaggle: Carvana Image Masking Challenge.
- Automating image segmentation provides a systematic way of segmenting an image quickly as soon as it is acquired which makes it useful in production environments where speed matters more than accuracy or precision alone.

## Summary

Reliable image segmentation is a crucial task in computer vision, particularly in medical imaging and self-driving vehicles. Manual segmentation of images is time-consuming and prone to errors, making automated treatment necessary for accuracy and efficiency. In recent years, various methods have been proposed to create CNNs that can produce a segmentation map for an entire input image in a single forward pass. One of the most successful deep learning methods is based on Fully Convolutional Networks (FCN), which uses CNN as a feature extractor by replacing fully connected layers with convolutional layers. Pixel-wise image segmentation is a demanding task in computer vision, especially for medical images and satellite images. Classical U-Net architectures composed of encoders and decoders are popular for such segmentations. Neural networks initialized with weights from pre-trained networks on large datasets like ImageNet show better performance than those trained from scratch on small datasets. In some practical applications, such as medicine and traffic safety, model accuracy is of utmost importance. In this paper, the authors present TernausNet: U-Net with VGG11 Encoder Pre-Trained on ImageNet for Image Segmentation. They demonstrate how the U-Net architecture can be improved by using the pre-trained encoder. The code and corresponding pre-trained weights are publicly available at https://github.com/ternaus/TernausNet. The authors compare three weight initialization schemes: LeCun uniform, the encoder with weights from VGG11, and full network trained on the Carvana dataset. This network architecture was part of the winning solution (1st out of 735) in the Kaggle: Carvana Image Masking Challenge. Every image consists of pixels that need to be divided into meaningful regions or classes for dense image segmentation. The most straightforward approach to this problem is manual segmentation; however, it is inefficient and prone to inaccuracies when human data curators are involved due to its laborious nature. Automating this process provides a systematic way of segmenting an image quickly as soon as it is acquired which makes it useful in production environments where speed matters more than accuracy or precision alone . For more details on these topics see [5].