ribin-baby / u2net_pytorch

pytorch implimentation of u2net architecture

U2 NET

Fig.1 - U2Net Architecture

• U2-Net is a two-level nested U-structure architecture. It uses a novel ReSidual U-block (RSU) module to extract multi-scale features without degrading resolution, allowing the network to go deeper and attain high resolution without significantly increasing memory and computation cost.
• used for for salient object detection, image segmentation, Image Matting, background removal and other image2image modeling tasks.

Fig.2 - UNet or RSU Block

• U-Net is a U-shaped encoder-decoder architecture with residual connections between each layers. It captures contextual information and intricate detail.

• These U-Net blocks in U2Net architecture are called ReSidual U-block or RSU.

• Example: we have trained an Image Matting model on P3M-10k dataset, and the results are given below.

Fig.3 - Image Matting with U2-Net example

Fig.4 - Image Matting with U2-Net training progress after each steps