🔩 DCNN-Based Screw Detection for Automated Disassembly Processes

This repository provides the complete source code and resources for the paper DCNN-Based Screw Detection for Automated Disassembly Processes.

Key Features:

• 🎯 6 Pre-trained CNN Models: Xception, InceptionV3, InceptionResNetV2, DenseNet201, ResNet101V2, ResNeXt101
• 🔬 Three Usage Modes: Classification only, Standalone detection, ROS integration
• 📊 98.7% Accuracy: State-of-the-art screw classification
• 🚀 Modern TensorFlow 2.x: Updated from legacy TF 1.x codebase
• 🐍 Pure Python: No ROS required for standalone usage

---

📋 Table of Contents

• Quick Start
• Installation
• Usage Modes
  • Mode 1: Classification Testing
  • Mode 2: Standalone Detection
  • Mode 3: ROS Integration
• Training Models
• Evaluation
• Project Structure
• Available Models
• Citation
• Contact

---

⚡ Quick Start

Test Pre-trained Model (No Dataset Required)

git clone https://github.com/eyildiz-ugoe/screw_detection.git
cd screw_detection
pip install -r requirements.txt

# Run detection evaluation on sample data
python evaluate/evaluate_detection.py \
  --det_path evaluate/det_2keras.txt \
  --gt_path evaluate/gt_test.txt \
  --no-plot

Expected output: AP = 0.1572

Full Detection on Your Own Image

# Download models first (see Installation section)
python scripts/standalone_detection.py \
  --image your_image.jpg \
  --model1 models/xception.h5 \
  --model2 models/inceptionv3.h5 \
  --output result.jpg

---

🔧 Installation

1. Clone Repository

git clone https://github.com/eyildiz-ugoe/screw_detection.git
cd screw_detection

2. Install Dependencies

pip install -r requirements.txt

Requirements:

• Python 3.7+
• TensorFlow 2.x (GPU support included)
• OpenCV
• NumPy
• Matplotlib (optional, for plotting)

Note: TensorFlow 2.20+ requires cuDNN 9.3.0+. If you have an older cuDNN (e.g., 8.9.7), either upgrade cuDNN or use CPU mode by setting CUDA_VISIBLE_DEVICES="".

3. Download Dataset and Pre-trained Models

Option A: Automated Download (requires unrar):

python scripts/download_data.py

Option B: Manual Download:

• Dataset: ScrewDTF dataset (~1.8GB)
• Pre-trained weights: Model weights on Google Drive (~2.1GB - 6 models)

Note: The dataset from Zenodo comes in TFRecord format. After extraction, run the conversion script to convert TFRecords to JPG images:

python scripts/convert_tfrecords.py

Expected folder structure after extraction and conversion:

screw_detection/
├── ScrewDTF/              # Dataset
│   ├── Train/
│   │   ├── label_0/       # Non-screw images
│   │   └── label_1/       # Screw images
│   ├── Eval/              # Validation set
│   └── Test/              # Test set
└── models/                # Pre-trained weights
    ├── xception.h5
    ├── inceptionv3.h5
    ├── inceptionResNetv2.h5
    ├── densenet201.h5
    ├── resnet101v2.h5
    └── resnext101.h5

---

🎯 Usage Modes

This repository supports three different usage modes depending on your needs:

Mode 1: Classification Testing

Test pre-trained classifiers on pre-cropped screw images.

Use when: Testing model accuracy, comparing architectures, benchmarking.

Example:

# Test Xception model
python src/prediction_xception.py \
  --weights models/xception.h5 \
  --screw-dir ScrewDTF/Test/label_1 \
  --non-screw-dir ScrewDTF/Test/label_0

# Output:
# ==============================
# TP: 1843 TN: 1974 FP: 8 FN: 42
# Accuracy: 0.9871

Test all 6 models at once:

python scripts/test_all_classifiers.py

Available architectures: xception, inceptionv3, inceptionresnetv2, densenet201, resnet101v2, resnext101

---

Mode 2: Standalone Detection 🆕

End-to-end screw detection on any image (no ROS required).

Use when: Processing static images, batch detection, development/debugging.

Pipeline:

1. Find circular candidates (Hough Transform)
2. Extract patches around candidates
3. Classify with CNN ensemble (Xception + InceptionV3)
4. Output bounding boxes with confidence scores

Basic usage:

python scripts/standalone_detection.py \
  --image your_image.jpg \
  --model1 models/xception.h5 \
  --model2 models/inceptionv3.h5 \
  --output result.jpg

Example output:

Finding circular candidates...
Found 3 circular candidates
Classifying candidates...
Detected 1 screws

Detected 1 screws:
  1. Position: (832, 82), Radius: 20, Confidence: 0.887

Advanced options:

# Adjust detection threshold
python scripts/standalone_detection.py --image input.jpg --threshold 0.3 --output result.jpg

# Tune Hough parameters for different screw sizes
python scripts/standalone_detection.py --image input.jpg \
  --min-radius 10 \
  --max-radius 50 \
  --hough-upper 150 \
  --hough-lower 30 \
  --output result.jpg

# Use different model combination
python scripts/standalone_detection.py --image input.jpg \
  --model1 models/resnet101v2.h5 \
  --model2 models/densenet201.h5 \
  --output result.jpg

# Display result window
python scripts/standalone_detection.py --image input.jpg --show

# Use GPU
python scripts/standalone_detection.py --image input.jpg --use-gpu --output result.jpg

Batch processing:

#!/bin/bash
for img in input_images/*.jpg; do
    output="results/$(basename $img)"
    python scripts/standalone_detection.py --image "$img" --output "$output"
done

---

Mode 3: ROS Integration 🤖

Full detection pipeline integrated with ROS (Robot Operating System).

Use when: Integrating with robotics hardware, real-time camera streams, ROS-based systems.

File: scripts/ros_detection.py (formerly candidate_generator.py)

Requirements:

• ROS (Kinetic/Melodic/Noetic)
• TensorFlow 1.x (for compatibility with original system)
• ROS cv_bridge, sensor_msgs

Features:

• Subscribes to ROS Image topics
• Real-time Hough circle detection
• CNN classification
• Publishes detection results to ROS topics
• Dynamic reconfigure for parameter tuning

Note: This mode requires a full ROS workspace setup and is intended for robotics integration. For standalone testing, use Mode 2 instead.

---

🎓 Training Models

Train your own classifiers on custom datasets:

# Train Xception (71×71 images)
python src/train_xception.py \
  --train-dir ScrewDTF/Train \
  --val-dir ScrewDTF/Eval \
  --output-weights models/my_xception.h5 \
  --epochs 16 \
  --batch-size 8

# Train InceptionV3 (139×139 images)
python src/train_inceptionv3.py \
  --train-dir ScrewDTF/Train \
  --val-dir ScrewDTF/Eval \
  --output-weights models/my_inceptionv3.h5

# View all training options
python src/train_xception.py --help

Training options:

• --train-dir: Training images organized by class (label_0/, label_1/)
• --val-dir: Validation images organized by class
• --output-weights: Where to save the best model
• --epochs: Number of training epochs (default: 16)
• --batch-size: Mini-batch size (default: 8)
• --learning-rate: Adam learning rate (default: 1e-4)
• --freeze-layers: Freeze N layers from backbone
• --initial-weights: Start from existing checkpoint

---

📊 Evaluation

Classifier Evaluation

Evaluate a trained model on test images:

python src/prediction_xception.py \
  --weights models/xception.h5 \
  --screw-dir ScrewDTF/Test/label_1 \
  --non-screw-dir ScrewDTF/Test/label_0 \
  --threshold 0.5

Output:

==============================
TP: 1843 TN: 1974 FP: 8 FN: 42
Accuracy: 0.9871

Detection Pipeline Evaluation

Reproduce the paper's precision/recall metrics:

# Evaluate Keras-based detector
python evaluate/evaluate_detection.py \
  --det_path evaluate/det_2keras.txt \
  --gt_path evaluate/gt_test.txt

# Evaluate TensorFlow-based detector  
python evaluate/evaluate_detection.py \
  --det_path evaluate/det_2tf.txt \
  --gt_path evaluate/gt_test.txt

What these files are:

• det_2keras.txt / det_2tf.txt: Pre-computed detection results from the paper
• gt_test.txt: Ground truth bounding box annotations
• Allows reproducing precision/recall curves without running full detection

---

🗂️ Project Structure

screw_detection/
│
├── README.md                           # This file
├── requirements.txt                    # Python dependencies
├── .gitignore                          # Git exclusions
│
├── scripts/                            # 🆕 Utility & runnable scripts
│   ├── download_data.py                # Download dataset & model weights
│   ├── convert_tfrecords.py            # Convert TFRecord dataset to JPG
│   ├── standalone_detection.py         # End-to-end detection (no ROS)
│   ├── ros_detection.py                # ROS-integrated detection
│   └── test_all_classifiers.py         # Test all 6 models
│
├── src/                                # Core library code
│   ├── pipelines/
│   │   ├── __init__.py
│   │   └── classification.py          # Training/inference logic
│   │
│   ├── resnet.py                       # ResNeXt implementation
│   │
│   ├── train_*.py                      # Training CLIs (6 architectures)
│   └── prediction_*.py                 # Inference CLIs (6 architectures)
│
├── evaluate/                           # Evaluation scripts & data
│   ├── evaluate_detection.py          # Precision/recall computation
│   ├── det_2keras.txt                 # Sample detection results
│   ├── det_2tf.txt                    # Sample detection results
│   └── gt_test.txt                    # Ground truth annotations
│
├── tests/                              # Unit tests
│   ├── conftest.py                    # Pytest import configuration
│   ├── test_evaluate_detection.py
│   └── test_pipelines_classification.py
│
├── assets/                             # Example output images
│   └── output.png
│
# Data folders (not in git, download separately):
├── models/                             # Pre-trained weights (.h5)
└── ScrewDTF/                           # Training/test dataset
    ├── Train/
    ├── Eval/
    └── Test/

---

🏆 Available Models

All models are trained on the ScrewDTF dataset and achieve high accuracy:

Model	Input Size	Weights Size	Test Accuracy	Use Case
Xception ⭐	71×71	80 MB	98.71%	Best overall, fast inference
ResNet101V2 ⭐	64×64	489 MB	97.60%	High accuracy, smallest input
ResNeXt101	64×64	485 MB	92.29%	Good accuracy
InceptionResNetV2	139×139	625 MB	91.18%	Larger input, more detail
InceptionV3	139×139	251 MB	91.08%	Good for ensemble
DenseNet201	221×221	212 MB	89.60%	Largest input size

Model weights: Download all 6 models from Google Drive

---

🧪 Testing

Run unit tests to verify installation:

# All tests
python -m pytest tests/ -v

# Specific test
python -m pytest tests/test_evaluate_detection.py -v

# Test all classifiers (requires dataset and models)
python test_all_classifiers.py

---

📚 Citation

If this work is helpful in your research, please cite:

@inproceedings{yildiz2019dcnn,
  title={DCNN-Based Screw Detection for Automated Disassembly Processes},
  author={Yildiz, Erenus and W{"o}rg{"o}tter, Florentin},
  booktitle={2019 15th International Conference on Signal-Image Technology \& Internet-Based Systems (SITIS)},
  pages={187--192},
  year={2019},
  organization={IEEE}
}

Paper: IEEE Xplore

Dataset: ScrewDTF on Zenodo

Model Weights: Pre-trained models on Google Drive (all 6 models)

---

💡 GPU vs CPU Usage

GPU Mode (default, requires CUDA/cuDNN):

python standalone_detection.py --image input.jpg --use-gpu --output result.jpg

CPU Mode (recommended if CUDA version mismatch):

# Method 1: Environment variable
CUDA_VISIBLE_DEVICES="" python standalone_detection.py --image input.jpg --output result.jpg

# Method 2: No GPU flag (default in most scripts)
python standalone_detection.py --image input.jpg --output result.jpg

Important: TensorFlow 2.20+ requires cuDNN 9.3.0+. If you have cuDNN 8.9.7 or encounter GPU errors, use CPU mode or downgrade TensorFlow: pip install tensorflow==2.15

---

📞 Contact & Support

• Issues: GitHub Issues
• Author: Erenus Yildiz

---

📝 License

See repository for license details.

---

🎉 What's New in This Version

This is a refactored version with significant improvements:

✅ TensorFlow 2.x Support: Modernized from legacy TF 1.x
✅ Standalone Detection: No ROS required for basic usage
✅ Unified API: Consistent interfaces across all architectures
✅ Better Documentation: Single comprehensive README
✅ Cleaner Structure: Organized files and clear naming
✅ Unit Tests: Automated testing for reliability
✅ ResNeXt101 Fix: Fixed import issues
✅ GPU/CPU Flexibility: Easy switching between modes

Legacy files (TF 1.x, kept for reference):

• evaluate/evaluate_classifiers_*.py - Original evaluation scripts (require TF 1.x)
• Kept for historical reference; not needed for normal usage

---