Mars Surface Segmentation with CBAM-UNet

A comprehensive semantic segmentation pipeline for Mars surface imagery using a Convolutional Block Attention Module (CBAM) – enhanced U‑Net architecture. This repository provides scripts for data preprocessing, model training, evaluation, inference, visualization, and pseudo‑labeling support.

🧠 Model & Methodology

Architecture: U‑Net with integrated CBAM (Channel + Spatial Attention) blocks

Segmentation Classes:

• 0 – Background (0, 0, 0)
• 1 – Crater (255, 255, 0)
• 2 – Rough (255, 0, 0)
• 3 – Smooth (0, 255, 0)
• 4 – Alluvial Fan(0, 0, 255)
• 5 – Boulders (139, 69, 19)

Loss Function: Weighted Cross‑Entropy + Dice Loss (background ignored)

Training: PyTorch with mixed‑precision (AMP) and OneCycleLR scheduler

Evaluation Metrics: Per-class IoU, Dice, Precision, Recall; Mean IoU; Average inference time

---

---

Contibutors: Anoushka Chatterjee and Keerthana A R