models

Modern models are based on U-Net like shape - using high receptive field by encoder decoder model with skip connection methods;

Resources

• Comparative Study of Image Segmentation Architectures Using Deep Learning - https://scribe.rip/3743875fd608

• Image Segmentation: Architectures, Losses, Datasets and Frameworks - https://neptune.ai/blog/image-segmentation

• Complete Guide to Semantic Segmentation - https://www.superannotate.com/blog/guide-to-semantic-segmentation

UpSample method

for Signal Processing

• insert zero

• apply low pass filter

One could generalize the model by using any given interpolation method instead of applying Low Pass Filter.

for Image Processing

• increase zero

• Apply Interpolation

upsample layer

• The conventional upsampling methods are not adaptive to the loss.

• use learned filter coefficients (LPF).

• The concept of Transposed Convolution was introduced in Fully Convolutional Networks for Semantic Segmentation.

• in pytorch:

  • https://pytorch.org/docs/stable/generated/torch.nn.ConvTranspose2d.html

• Adjoint Operator

Resources

• What is Transposed Convolutional Layer - https://scribe.rip/40e5e6e31c11

• Understand Transposed Convolutions - https://scribe.rip/4f5d97b2967

• Dive into Deep Learning - Computer Vision - Transposed Convolution - https://d2l.ai/chapter_computer-vision/transposed-conv.html