Real-Time Smoke Detection Using Hybrid Image Processing Model

Overview

This project aims to develop a real-time smoke detection system using a hybrid image processing model that combines HSV thresholding and a custom Pixel CNN. This approach is designed to enhance the accuracy and efficiency of detecting smoke in various scenarios, including 3D data such as medical scans and videos.

Methods

Method 1: HSV_Threshold Image Processing

• Description: A no-machine learning model suitable for real-time detection.
• Performance: Surpassed expectations in detecting no-smoke images during testing.

Method 2: Custom based Pixel CNN Model

• Description: Useful for processing 3D data like medical scans and videos due to its ability to analyze spatial relationships in all directions.

Method 3: Hybrid Model (Final Model)

• Description: Combines HSV thresholding and Pixel CNN for better accuracy.
• Advantages: As HSV is a no-CNN approach, training and testing time is avoided, making it advantageous for real-time detection.

Pre-Processing

• Tools: Utilized OpenCV and Pillow libraries.
• Techniques: Thresholding, contour analysis, and Hough Transform for circle detection.
• Frame Size: Resized frames to 244x244 for optimal processing.
• Noise Reduction: Applied Gaussian Blur to remove noise.

Training, Validation & Testing

• Dataset: 60 videos.
  • Training: 80% (48 videos)
  • Validation: 10% (6 videos)
  • Testing: 10% (6 videos)
• Parameters:
  • Number of epochs: 10
  • Batch size: 16
  • Training time: 2 hours for input size 244x244, 48 videos, 10 epochs, batch size 16.
• Hardware:
  • Platform: Kaggle private virtual machines.
  • GPU: T4 x2
  • Tools: TensorFlow 2.15.1, Keras, OpenCV

Architecture

Metrics

• HSV-Threshold:
  • Accuracy: 89.10%
  • F1-Score: 50.22
  • Recall: 33.98
  • Precision: 59.56
• Custom Pixel CNN:
  • Accuracy: 86.83%
  • F1-Score: 55.97
  • Recall: 62.12
  • Precision: 50.93
• Hybrid:
  • Accuracy: 83.13%
  • F1-Score: 82.9
  • Recall: 83.16
  • Precision: 82.7

Inference Speed

• HSV-Threshold: 99 seconds
• Custom Pixel CNN: 1066 seconds
• Hybrid: 816 seconds

Inference Results

Real-Time Smoke Detection

• Process:
  • Videos are processed 10 frames at a time to consider the temporal correlation of sequences.
  • Predictions are made using the hybrid model for smoke or no smoke.
  • Frames are labeled and stitched back together for real-time output.

Conclusion

• The hybrid model provides a robust and efficient solution for real-time smoke detection, leveraging the strengths of both HSV thresholding and custom Pixel CNN technique.