The introduction of U-net

The introduction of U-net

The U-net is a kind of deep neural networks and first presented in paper U-Net: Convolutional Networks for Biomedical Image Segmentation. And now it is generally used for the image segmentation task and object detection task.

Since 2012, deep learning has achieved great success in image classification. And a lot of great network have been presented, Such as Alexnet, which is seen the first deep neural networks, VGG , inception net, and Resnet, for the first time, surpassing humans in image classification tasks.

Motivated by the success of image classification based on convolutional neural
networks(CNNs), a lot of researchers also tried to use CNNs to improve image segmentation tasks. However, the results were not as good as image classification.

The difference between those tasks is that the output of image classification is a label, which only need the fine features (the deeper layers of CNNs). But the desired output of image segmentation and object detection should include localization and a class label that is supposed to be assigned to each pixel, which the coarse features(the shallow layers of CNNs),for example, edges, are needed.

The tremendous things of the paper is that they presented a novel architecture that use skip connection (looks like Resnet in some way) to use both coarse and fine features, which different from the tradition CNNs only including down-sampling units.

After this paper, some researchers issued some improved U-net, for example using dense-net instead of Resnet and so on. And the better and better results are come true in image segmentation and object detection, which provide much help to auto-driving filed.

The introduction of U-net_第1张图片

Reference

[1] : https://arxiv.org/pdf/1505.04597.pdf

你可能感兴趣的:(Deeplearning,U-net)