Skip to content

README.md

Latest commit

 

History

History
117 lines (74 loc) · 4.93 KB

README.md

File metadata and controls

117 lines (74 loc) · 4.93 KB

🧠 Gabor Wavelet-based Image Classification using Convolutional Neural Networks (CNNs)

This repository presents a deep learning-based approach to improve image classification performance by leveraging Gabor wavelet representations of images. The core idea is to train multiple CNN models using both original and Gabor-transformed images and perform ensemble decision fusion to enhance classification accuracy.

📌 Project Summary

We apply this approach to classify chest X-ray images for pneumonia detection, combining traditional CNN architectures with directionally localized features extracted via Gabor filters.

🔍 Motivation

  • Improving diagnostic tools for pneumonia detection using X-rays.
  • Exploiting the spatial-frequency sensitivity of Gabor wavelets.
  • Combining signal processing and deep learning to improve generalization.

🌐 Methodology

🌀 Gabor Wavelets & Filters

  • Gabor wavelets are complex exponentials modulated by Gaussians, sensitive to frequency and orientation.
  • Gabor filters act as localized, orientation-tuned edge detectors.

🔄 Image Transformation Pipeline

  • For each input image:
    1. Apply Gabor filters at 8 different orientations.
    2. Generate 8 Gabor-transformed images.
    3. Train 9 CNNs (1 with the original image + 8 with Gabor-transformed variants).

🧮 Decision Fusion

  • Final prediction is made by aggregating the output probabilities from all CNNs using the Sum of Probabilities (SMP) Rule:
  • A custom threshold of 0.4 is used for final classification.

📊 Dataset

  • Source: Pediatric chest X-ray dataset from Guangzhou Women and Children’s Medical Center.
  • Size: 5,856 JPEG images
    • Training: 5,216
    • Validation: 16
    • Testing: 624
  • Classes: Pneumonia, Normal

🏗️ CNN Architecture

Standard convolutional architecture with ReLU activations and max-pooling. Each CNN is trained independently on either the raw image or a specific Gabor-transformed version.

🧠 Trained Models

The repository includes a models/ folder, that contains the 9 trained CNN models used in this project. These models are saved as final_model_0.h5 to final_model_9.h5.

Each model corresponds to a CNN trained on a specific representation of the input image:

  • final_model_0.h5 is trained on the original image (no Gabor transformation).
  • final_model_1.h5 to final_model_8.h5 are trained on Gabor-transformed images, each filtered at a specific orientation.

📐 Orientation Mapping

Model File Description Gabor Filter Orientation
final_model_0.h5 Trained on original input image None
final_model_1.h5 Trained on Gabor-transformed image 0°
final_model_2.h5 Trained on Gabor-transformed image 22.5°
final_model_3.h5 Trained on Gabor-transformed image 45°
final_model_4.h5 Trained on Gabor-transformed image 67.5°
final_model_5.h5 Trained on Gabor-transformed image 90°
final_model_6.h5 Trained on Gabor-transformed image 112.5°
final_model_7.h5 Trained on Gabor-transformed image 135°
final_model_8.h5 Trained on Gabor-transformed image 157.5°

These models are used during inference for ensemble prediction via decision fusion, aggregating the outputs of all 9 models to produce the final classification result.

📈 Results

Model Accuracy Precision Recall F1-Score
Single CNN (Raw) 90.06% 92.27% 91.79% 0.92
Ensemble (Fused) 92.47% 90.73% 97.94% 0.94

The ensemble model significantly improves recall while maintaining high precision.

🔭 Future Work

  • Try alternative wavelet families for image decomposition.
  • Explore other CNN variants (e.g., ResNet, EfficientNet).
  • Investigate wavelet packet decomposition for richer representations.

📚 References

  1. Granlund, G. H. (1978). In Search of a General Picture Processing Operator. Computer Graphics and Image Processing, 8(2), 155–173. https://doi.org/10.1016/0146-664X(78)90047-3

  2. Serte, S., & Demirel, H. (2019). Gabor wavelet-based deep learning for skin lesion classification. Computers in Biology and Medicine, 113, 103423. https://doi.org/10.1016/j.compbiomed.2019.103423

👨‍🎓 Author

Rajdeep Pathak
Department of Mathematics,
Indian Institute of Technology Hyderabad
📅 May 6, 2025