lz666win / deeplabv3plus-keras

deeplabv3plus (Google's new algorithm for semantic segmentation) in keras：Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation

Deeplab v3+ in keras

DeepLab is a state-of-art deep learning model for semantic image segmentation. For tensorflow implemention, please refer to [tensorflow/models/research]: https://github.com/tensorflow/models/tree/master/research/deeplab

The Features of DeepLab v3+

• Encoder-Decoder Structures
• Atrous Convolutions in ASPP structure
• Separable Convolutions instead of traditional Convolutions
• Bilinear Interpolation instead of Transpose Convolutions

The Brief Structure

Encoder

1. Xception as the Feature Extractor
2. ASPP
3. Separable Convolutions

Decoder

1. Upsampling
2. Separable Convolutions