Research
Image Analysis
Noise-enhanced Edge-preserving Image Denoising
Edge-preserving denoising using gradient-based estimation and iterative noise-aided processing
In this work, the concept of stochastic resonance has been applied to denoise the noisy input image. In stochastic resonance, a small value of noise is added to the noisy input image iteratively to enhance the edges of the input image based on which the weight coefficient is determined. It is observed the quality of the output image improves as the number of iterations is increased up to a certain number of iteration after which the quality of the image degrades or remains constant depending on the value of noise in the input image. In order to compare the quality of the denoised output image wrt the number of iterations state of the art no-reference quality metric such as Blind Universal Quality Indices (BIQI), spatial-spectral entropy-based quality (SSEQ) have been used. When compared with conventional denoising techniques such as Anisotropic diffusion (AD), Bilateral filtering (BLT), adaptive smoothing, the proposed algorithm yields marginally better results.
Noise-aided edge-preserving image denoising using non-local means with stochastic resonance
In this paper, we extend the application of stochastic resonance to improve the performance of the conventional non-local means (NLM) filtering for edge-preserving image denoising. The NLM algorithm typically involves computation of weights denoting similarity of a pixel with all other pixels in the image. In the proposed algorithm, these similarity weights are iteratively processed using the concept of dynamic stochastic resonance. The results indicate a significant improvement in sharpness of edges in the denoised images in comparison with the conventional NLM approach both visually and quantitatively in terms of full-reference and no-reference image quality metrics.
Enhanced edge detection using SR-guided threshold maneuvering and window mapping
Preserving edges in a noisy environment is a challenging task as even some of the latest end-to-end deep learning (DL) algorithms continue to struggle in achieving high pixel-level accuracy. As the Canny Edge Detector (CED) continues to be one of the most popular edge detection operators, this paper presents an enhanced CED using Stochastic Resonance (SR) guided threshold maneuvering and window mapping, which takes the same input parameter set as that of the conventional Canny but produces the edge map with better-connected edges and reduced noise. The SR-based analysis informs the steps that should be followed to enhance the performance of the classical CED. We also propose a new measure for efficient edge detection; a unique, efficient way of edge content extraction and its combination for various channels; and a framework to handle repercussions of the randomness of the noise. Since the proposed solution comes in the form of a modular patch-based framework, it can be easily incorporated into other algorithm developments. Qualitative and quantitative results are presented along with the BSDS500 & BIPED benchmarking to showcase the proposed algorithm’s effectiveness. On BIPED benchmarking, our algorithm gives the human-level performance (F1 score .79), which is appreciable considering that it is a non-DL–based algorithm.
Video_Access-2021-42494.mp4Image Quality Assessment
No-reference Image Quality Assessment using Gradient-based Structural Integrity and Latent Noise Estimation
Image quality assessment (IQA) plays a crucial role in monitoring quality control in image communication systems, and in benchmarking and optimizing parameters in enhancement algorithms. The full-reference IQA metrics require a good-quality reference image, obtaining which may not be practical in real-life applications. This paper, therefore, proposes a no-reference IQA metric based on the hypothesis that every image has latent additive white Gaussian noise (AWGN). A mathematical model was developed on a dataset of fifty test images by computing gradient-based structural similarity of corrupted images w.r.t. the original. Statistical modeling of the observations were found to fit an exponential parametric model. The standard deviation of the latent (or apparent) AWGN present in any image was estimated using an SVD-based approach. The proposed metric, referred to as the no-reference gradient-based structural integrity (NRGSI), is then computed by a simple backprojection of the estimated noise deviation on the exponential model. The accuracy of the proposed objective metric is characterized by its comparison with subjective quality scores given by ten subjects, and with a classical perceptual quality measure.
No-reference image quality assessment using Gabor-based smoothness and latent noise estimation
No-reference image quality assessment is a challenging task due to the absence of a reference image in practical situations to quantify image quality. This paper proposes a new no-reference image quality metric for natural images using latent noise estimation, Gabor response, and contrast deviation. The algorithm employs an extension of gradient-based SSIM into the no-reference application using SVD-based AWGN estimation, and defines attributes such as Gabor-based smoothness and contrast deviation. The proposed metric arrives at an overall quality score by computing a linear weighted summation of the three image attributes. The proposed algorithm has been tested on several public databases (i.e. LIVE, TID 2013 and CSIQ), and the overall results display a noteworthy correlation of nearly 80% with the human visual system.
Quantifying image naturalness using Curvelet Features
Distinguishing computer-generated (CG) images from natural images is an effortless job for the human eye but not for a machine. Artificial or CG image processing services are growing rapidly due to popular smartphone applications and filters in social media applications. The traditional image quality assessment (IQA) metrics are mostly defined for real-world images in terms of attributes describing noises and distortions. In contrast with traditional natural image content, artificial or CG image content has special characteristics that differentiate them from natural images. This difference opens new opportunities of research towards designing metrics that define ‘naturalness’ in terms of image attributes. In this work, we investigate how curvelet features of a natural image can represent naturalness of the image. By training various classifiers using differential curvelet features of artwork-like images, we report a novel approach to estimate image naturalness and its performance for various levels of naturalness. The reported results are promising and show potential of improvement using detailed feature engineering.
P. Shabari Nath, Harsh K. Gandhi, and R. Chouhan, "Quantifying image naturalness using differential curvelet features," Proc. Image and Vision Computing New Zealand, 09 - 10 December 2021.