DOI: 10.46810/tdfd.1872328 ISSN: 2149-6366

IGHF-DL: PARAMETER-EFFICIENT HYBRID IMAGE DENOISING USING CLASSICAL MATHEMATICAL FILTERING AND DEEP LEARNING

Fatih Maraşlı, Serkan Öztürk
Show PDF Cite
There is a critical balance between denoising and detail preservation that requires careful consideration. There is also a need for mathematically grounded, computationally efficient denoising models that are practical for real-world applications. In this study, we propose the IGHF-DL model, which improves the mathematical edge preservation power of the classic Inverse Gaussian Harmonic Filter (IGHF) by incorporating a CNN structure that includes six residual blocks, a channel attention mechanism, and an edge-sensitive enhancement block. In our model, training was performed with a low computational cost, reducing the number of DnCNN parameters by 54%. In comprehensive experiments performed on the BSD68, CBSD68, McMaster, Kodak24, Set12, and Urban100 datasets, our model achieved a PSNR value of 28.72 dB for BSD68, outperforming the BM3D (+0.15 dB) model and reaching 98.3% of the DnCNN performance, demonstrating competitive performance. Our model is successful not only in terms of metrics but also in terms of edge preservation and perceptual quality. On datasets with textured images such as Urban100, IGHF-DL outperforms the IGHF model in edge preservation and perceptual quality with FOM: 0.8252 and LPIPS: 0.1147, demonstrating its robustness on a mathematical basis and showing potential for further development in integration with next-generation methods. At the same time, compared to the IGHF model, IGHF-DL showed the best improvement in metrics at high noise levels with a +7.44 dB improvement. The proposed hybrid approach offers a practical and efficient solution with parameter efficiency that reduces computational costs for resource-constrained environments.

More from our Archive