Personal project

Pavement Crack Segmentation

Fine-grained pavement crack segmentation using DAResUNet-ASPP: a Dual-Attention Residual U-Net with multi-scale context and boundary supervision.

Computer Vision ResearchDeep learning project

Explore Visual segmentation comparison

Three pavement crack samples comparing input images, ground truth, U-Net and proposed masks, green overlays, and error maps with red false positives, blue false negatives, and white true positives

Tech Stack

PythonPyTorchOpenCVNumPyMatplotlibGoogle ColabKaggle

Feature spotlights / Pavement Crack Segmentation

Visual segmentation comparison

This qualitative comparison from the seed 42 experiment shows where predicted crack regions align with the reference masks. The shared sample layout supports direct visual inspection of both models.

Visual segmentation comparison

Predicted masks, overlays, and errors

Three fixed test samples place the input image and ground truth beside U-Net and the proposed model. Each model includes a predicted mask, an overlay, and an error map at threshold 0.5: red marks false positives, blue false negatives, and white true positives.

Inspect crack boundaries and prediction errors on the same samples.

Challenge

Fine cracks on complex pavement surfaces require detailed feature extraction across patterns and scales.

Represent thin crack boundaries and varied pavement patterns while evaluating each model component.

Approach

Combined residual learning, attention, ASPP, and boundary supervision, with ablations, multiple seeds, and threshold analysis.

Explores model architecture and evaluation methods for fine-grained pavement crack detection.

Implementation

Residual U-Net encoder-decoder
Dual attention and attention gates
Compact ASPP for multi-scale context
Strip pooling
Deep supervision
Edge and boundary supervision

Features

Fine-grained crack segmentation
Crack500, CrackForest, DeepCrack, and non-crack images
Dataset of 3,548 images
LOCO ablation study
Multi-seed validation
Threshold sensitivity analysis

Highlights

DAResUNet-ASPP architecture for pavement segmentation
Dice, IoU, accuracy, precision, recall, F1, and pixel-wise AP evaluation
GPU profiling with NVIDIA Tesla T4

Let's talk about your next project.

Let's Talk

Loading page...

Back to projects

Personal project

Pavement Crack Segmentation

Fine-grained pavement crack segmentation using DAResUNet-ASPP: a Dual-Attention Residual U-Net with multi-scale context and boundary supervision.

Computer Vision ResearchDeep learning project

Explore Visual segmentation comparison

Tech Stack

PythonPyTorchOpenCVNumPyMatplotlibGoogle ColabKaggle

Feature spotlights / Pavement Crack Segmentation

Visual segmentation comparison

This qualitative comparison from the seed 42 experiment shows where predicted crack regions align with the reference masks. The shared sample layout supports direct visual inspection of both models.

Visual segmentation comparison

Predicted masks, overlays, and errors

Inspect crack boundaries and prediction errors on the same samples.

Challenge

Fine cracks on complex pavement surfaces require detailed feature extraction across patterns and scales.

Represent thin crack boundaries and varied pavement patterns while evaluating each model component.

Approach

Combined residual learning, attention, ASPP, and boundary supervision, with ablations, multiple seeds, and threshold analysis.

Explores model architecture and evaluation methods for fine-grained pavement crack detection.

Implementation

Residual U-Net encoder-decoder
Dual attention and attention gates
Compact ASPP for multi-scale context
Strip pooling
Deep supervision
Edge and boundary supervision

Features

Fine-grained crack segmentation
Crack500, CrackForest, DeepCrack, and non-crack images
Dataset of 3,548 images
LOCO ablation study
Multi-seed validation
Threshold sensitivity analysis

Highlights

DAResUNet-ASPP architecture for pavement segmentation
Dice, IoU, accuracy, precision, recall, F1, and pixel-wise AP evaluation
GPU profiling with NVIDIA Tesla T4

Let's talk about your next project.

Let's Talk