Floorplan Denoising & Segmentation

AttResUNet-based denoising model removes noise (text, dimensions, hatches) from architectural floorplans, enabling more accurate wall/door/window segmentation for BIM reconstruction.

---

1. Project Overview

This project, developed by the Construction AI & Robotics Team, proposes a Residual-Attention U-Net (AttResUNet) for architectural drawing denoising to enhance Mask R-CNN instance segmentation on floorplans.

• Dataset:

  • AI-Hub 2022 (48,033 images, 2.6M+ structure labels)
  • Type: Apartment / Multiplex / Detached house floorplans
  • Class: STR (Structure) / SPA (Space) / OBJ (Object) / OCR (Text)

• Performance:

  • Noise ROI: PSNR 35.43, SSIM 0.9978
  • Structure ROI: PSNR 24.14, SSIM 0.9852
  • Using class-weighted loss (background: 0.1 | wall: 0.3 | window/door: 1.0)

---

2. Key Results

Denoising Results:

Metric	Full Image	Noise ROI	Structure ROI
MSE	0.0046	0.0034	0.0608
PSNR	23.64	35.43	24.14
SSIM	0.9503	0.9978	0.9852
LPIPS	0.0629	0.0026	0.0361

Segmentation Results (mAP50):

Dataset	mAP50
Raw / Raw	0.2090
Raw / Clean	0.1032
Clean / Clean	0.0598
GT / Clean	0.1141

• The Raw / Raw configuration achieved the highest mAP50.
• However, GT / Clean outperformed Raw / Clean and Clean / Clean,
  showing that well-aligned denoised images and labels could further improve segmentation quality.

---

3. Denoising Model Architecture

Residual-Attention U-Net

• Residual blocks for stable deep feature learning and structure preservation
• Attention gates to focus on thin structural lines
• Hybrid encoder–decoder architecture specialized for architectural patterns

Loss Function

• BCE (0.5) + Dice + SSIM
  • Binary cross-entropy for mask binarization
  • Dice for overlapping region accuracy
  • SSIM for perceptual similarity
  • Class-weighted loss applied: background: 0.1 | wall: 0.3 | window/door: 1.0
    • Class weights were applied using the Mask dataset, assigning higher importance to structural regions such as walls, windows, and doors.

Training Settings

• Dataset split: Train 3,000 / Validation 300 / Test 300
• Epochs: 200 (early stopping enabled)
• Augmentation:
  • Flip (50%), translate ±10px, scale ±10%, rotation ±15° (70% prob.)

---

4. Dataset

• Source: AI-Hub Construction Drawing Dataset (2022)

  • Provided Formats: PNG images and JSON labels (structure, object, space annotations)

• Extended for Denoising:

  • Additional data created in this work:
    • Raw — noisy input floorplans
    • GT (Ground Truth) — clean reference drawings
    • Masks — region maps for wall, window, and door
  • These splits were derived from the AI-Hub dataset and refined for noise removal and structure segmentation tasks.

• Characteristics of Drawing Data:

• Binary-like high-contrast composition (0 and 1).

• Extremely sensitive to pixel-level variations — even single-pixel noise significantly affects object detection and segmentation accuracy.

---

5. Experiments

• Denoising Model:
  • Residual-Attention U-Net
• Segmentation Model:
  • Mask R-CNN

5.1 Denoising Evaluation

Evaluation Regions

• Full Image: Entire image
• Noise ROI: Regions containing text, dimensions, and hatch patterns
• Structure ROI: Regions with architectural elements (walls, doors, windows)

Quantitative Results

Metric	Full Image	Noise ROI	Structure ROI
PSNR	23.64	35.43	24.14
SSIM	0.9503	0.9978	0.9852
LPIPS	0.0629	0.0026	0.0361

Denoising performance is outstanding in noise-dominant regions,
but due to the high-contrast characteristics of floorplan data (binary-like pixel intensity),
the Structure ROI scores are relatively lower, reflecting the model’s sensitivity to fine structural edges.

---

5.2 Segmentation Tests

Segmentation Test Variants (Train Dataset / Test Dataset):

1. Raw / Raw
2. Raw / Clean
3. Clean / Clean
4. GT / Clean

Dataset	mAP50
Raw / Raw	0.2090
Raw / Clean	0.1032
Clean / Clean	0.0598
GT / Clean	0.1141

Tests (2) and (3) were designed with the expectation that denoised images would yield higher segmentation performance.
However, (1) Raw / Raw achieved the highest mAP50, while (2) and (3) underperformed due to label alignment with raw images, causing mismatches on clean inputs.
The (4) GT / Clean test showed noticeable improvement over (2) and (3), indicating that with properly aligned denoised images and labels, segmentation performance could be further enhanced.

Insights:

• Structure restoration > generic noise removal
• Proper label–image alignment is critical for segmentation performance
• Future directions : line continuity and junction-aware loss functions

---

6. Environment

Trained on Mac M4 and Google Colab environments.

---

7. Future Work

• Structure-aware denoising focused on line continuity and geometric accuracy
• Junction-aware loss for boundary refinement
• Integration into a real-time BIM reconstruction pipeline