Model Card for ResNet50 Cracks Finetuned
Model Details
- Model Name: ResNet50 Cracks Finetuned
- Base Architecture: ResNet50
- Task: Image Classification / Crack Analysis
- Framework: TensorFlow / Keras
Intended Use
This model is intended to analyze images of concrete cracks to determine severity. Working in conjunction with computer vision pipelines (e.g., Canny edge detection, Hough lines), the overall system evaluates the crack's attributes to predict structural risk.
Training Data
The dataset was processed with a custom augmentation and preprocessing pipeline. For example, before use, positive crack images were processed heavily including:
- Grayscale conversion
- Gaussian Blur
- Canny Edge detection
- K-Means Clustering on ResNet50 extracted features (using PCA for dimensionality reduction) to separate clusters of diverse crack distributions.
Metrics
- Features extracted were reduced to 100 components using PCA.
- Clustered using MiniBatchKMeans (K=3) offering a Silhouette Score of ~0.30, indicating overlapping yet distinct crack typologies.
Limitations and Biases
- The system heavily relies on proper preprocessing. Blur and edge parameters may need fine-tuning if the camera quality significantly varies.
- Training data is limited to specific types of concrete walls. Cracks on varied textures like painted walls or asphalt might yield unpredictable results.
