LECTURE 25

Summary

This lecture introduces complex-valued deep learning for SAR image analysis, focusing on how neural networks can use phase information in complex SAR data rather than discarding it.

Why use complex networks?

SAR data contains both magnitude and phase, and phase can carry important information about the observed scene. Conventional computer-vision architectures are designed for real-valued optical images, so they do not directly handle complex inputs. Complex-valued networks provide a way to work with complex SAR data and may also be useful for polarimetric SAR.

The lecture notes that phase may not benefit every task, but it can be valuable in many applications.

Building blocks of a complex CNN

The lecturer describes complex-valued counterparts to familiar neural-network components.

• Complex convolution: Let a filter be (W = A + iB) and an input be (H = X + iY). Their convolution has:
  • Real part: (A * X - B * Y)
  • Imaginary part: (B * X + A * Y)

Each output component therefore combines information from the real and imaginary parts of both the filter and the input.

• Complex activations: A standard real-valued ReLU cannot be applied directly to a complex number, so alternatives are used:

  • CReLU applies ReLU separately to the real and imaginary components.
  • ZReLU uses the input’s phase to decide whether to keep the complex value or set it to zero.
  • Other complex activation variants are also available.

• Other layers: Components such as batch normalization and weight initialization can also be adapted for complex networks.

Network architecture and training

Complex layers can be assembled into familiar architectures, including residual blocks with complex convolutions, complex batch normalization, complex activations, and skip connections.

Training requires suitable complex-valued objectives. The lecture gives an example of a complex average cross-entropy approach that evaluates the real and imaginary predictions against the same ground-truth labels. The appropriate loss and layer choices depend on the task, such as classification or regression.

Course context

This lecture concludes the week’s discussion of SAR image analysis. Earlier topics included object detection, despeckling, CycleGAN, knowledge distillation, handling different viewing angles, and working with different polarizations.

SAR and multisensor inputs will be revisited later in the course, including in discussions of change detection and foundation models.

Speakers and sources

• Speaker: Unnamed course lecturer.
• Other sources or speakers: None explicitly identified in the subtitles.