Upsampling Methods
- Nearest Neighbour: Simple duplication of pixel values.
- Bilinear/Linear Upsampling: Using weighted averages of surrounding pixels for smoother results.
- Transposed Convolutions: A learnable upscaling method.
UNet Architecture
- An “Encoder-Decoder” structure with a “contracting path”(downsampling) and an “expansive path”(upsampling).
- Skip Connections in UNet: Feature maps from the encoder are concatenated with the decoder maps to preserve high-resolution spatial details.
Key UNet Properties
- Fully Convolutional: It has no fully connected layers, meaning it can be trained and tested on any image size (provided it fits in memory). ◦ Receptive Field: Every output pixel is influenced by a specific area of the input. More levels in the UNet lead to a larger receptive field. ◦ Padding: Original UNets did not use padding (images got smaller), but modern versions often use “SAME” padding