A robust deep learning approach for impulse noise filtering using hybrid auto-encoder with fuzzy median filter

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Impulse noise degrades image quality, necessitating advanced removal techniques.
• De-noising convolutional neural networks (DnCNNs) show promise for image de-noising but require careful parameter selection.
• Existing methods struggle with high-density impulse noise, impacting image restoration effectiveness.

Purpose of the Study:

• To propose an effective image restoration technique for detecting and eliminating high-density impulse noise.
• To integrate DnCNNs with autoencoders and fuzzy median filters for enhanced de-noising.
• To evaluate the performance of the proposed deep learning de-noising technique using various metrics.

Main Methods:

• A novel technique integrating DnCNNs, autoencoders, and fuzzy median filters was developed.
• The deep learning model classifies noisy and clean pixels.
• The autoencoder with a fuzzy median filter reconstructs the de-noised image.

Main Results:

• The proposed method achieved high accuracy in classifying noisy and clean pixels.
• Quantitative metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) showed significant improvements.
• Experimental results demonstrated superior performance compared to conventional impulse noise filtering techniques.

Conclusions:

• The developed deep learning-based de-noising technique effectively removes high-density impulse noise.
• The integration of DnCNNs, autoencoders, and fuzzy median filters enhances image restoration quality.
• The method successfully preserves essential image details and features during the de-noising process.