Edge-Attentive Residual CNN
A residual CNN with attention, designed from scratch at Havelsan Technology Radar to reduce noise in radar imagery while preserving the structural detail that noise-reduction models typically blur away.
Problem
Denoising raw radar imagery without losing structural (edge) detail — generic denoising tends to smooth away exactly the fine structure that matters most for downstream analysis.
Flow
An end-to-end training pipeline built around a custom residual, attention-based architecture and an ensemble of models.
- 01Raw radar image input.
- 02A residual CNN with attention, designed from scratch.
- 03Hybrid activation and loss functions preserve high-frequency structural detail.
- 04Ensemble learning across multiple architectures.
- 05End-to-end training pipeline, validated against accuracy metrics.
Technologies
Notable decisions
Attention over a plain residual baseline
Attention layers were added on top of a residual backbone specifically to protect edge and structural detail, which a plain residual denoiser tends to smooth away along with the noise.
Hybrid activation and loss functions
Standard denoising losses optimize for smoothness; the loss and activation functions here were combined specifically to keep high-frequency structural information intact.
Ensemble across multiple architectures
Rather than relying on a single trained model, predictions are combined across multiple architectures to improve robustness on the final validation metrics.
Result
Built at Havelsan Technology Radar; reduces noise in radar imagery while preserving structural detail.