Steel-Defect-Detection

Overview This project is an end-to-end steel defect detection system leveraging a pre-trained ResNet152V2 deep learning model. The system identifies different types of defects on steel surfaces, providing bounding boxes and defect classifications.

It performs two main tasks:

1. Bounding Box Prediction: Detects the location of defects on steel surfaces.
2. Defect Classification: Classifies the type of defect from multiple categories.

The model is trained using a custom dataset with annotated bounding boxes and defect classes. The dataset includes images and corresponding XML annotation files.

Project Structure

• images/: Directory containing the steel defect images linked to download.
• label/: Directory containing XML annotation files with bounding box and class information Linked to download.
• resnet152v2.h5: The trained model (not included in this repository).
• steel_defect_detection.py: Jupyter notebook for training and saving the trained model.
• link.txt: Link to Python script to load the model and perform inference.

How to Use

1. Clone the repository:

   git clone https://github.com/MohiniDeshpande/steel-defect-detection.git

   cd steel-defect-detection

   pip install -r requirements.txt

2. Train the model or run inference:

To train the model, download data from links above and use the 'steel_defect_detection.py' file.

To run inference, use the colab link.

Model Architecture

The model is based on the ResNet152V2 architecture and is pre-trained on ImageNet. The model has been adapted for object detection (bounding box prediction) and classification of defects.

Layers: Base Model: Pre-trained ResNet152V2 (Frozen for initial training).

Bounding Box Prediction: Dense layers predicting xmin, ymin, xmax, ymax.

Defect Classification: Dense layers predicting the defect type.

Model Performance

The model achieved high accuracy on defect classification and precise bounding box predictions for defect locations.