Modified Euler Frobenius Halfband Polynomial Filter based Dual-Tree Complex Wavelet Transform and Metaheuristic Optimization Algorithms for Image Denoising along with Artificial Neural Networks

Abstract

In recent years Dual Tree Complex Wavelet Transform (DT-CWT) has been applied to heterogeneous image denoising problems. The quality of reconstructed images depends on low pass and high pass filter coefficients of the transform. The design of bi-orthogonal low pass filters and selection of an optimal set of filter coefficients is therefore necessary for effective image restoration. Hence the main contribution of our  paper is a DT-CWT based image denoising model, where low pass filters are designed using Modified Euler Frobenius Half Band Polynomial (MEFHBP) and filter coefficients are optimized by hybrid meta heuristic algorithms termed Earth Worm based Grey Wolf (EWGW) optimization. In the proposed de-noising process, DT-CWT of noisy image is initially found by employing appropriate MEFP based filters. The coefficients of filters are then optimized by EWGW algorithm, objective function being maximization of Peak Signal to Noise Ratio (PSNR). With optimized set of filter coefficients, final DT-CWT is performed resulting in denoised image. As an extension, ability of EWGW algorithm in optimizing weights of Artificial Neural Network (ANN) is also studied by means of classification model. In classification stage, noisy images first undergo multi-level thresholding based segmentation. Noise Power Spectrum and Bark frequency features are then extracted from the image and are fed to ANN whose weights are optimized by EWGW. Experiments are conducted on texture, medical and satellite image datasets. The performance of the algorithm is measured by means of various performance metrics and is compared with other optimization techniques. Finally statistical analysis is done to substantiate experimental results. The analysis shows that MEFHBP-EWGW outperforms many optimization techniques and results in better denoised images. It is also able to optimize ANN weights in a subtle manner.