Authors: Aishwarya Jagadish
Certificate: View Certificate
This research paper presents a comprehensive investigation into the realm of image enhancement techniques, spanning a spectrum of methodologies, from classical filtering to cutting-edge deep learning algorithms. The overarching objective of these techniques is to elevate image quality, sharpen details, and enhance clarity, with applications extending across various domains. Furthermore, this work underscores the ethical dimensions inherent to image enhancement in an era increasingly concerned with issues of privacy and authenticity.
In our increasingly visual world, pictures are more than just snapshots. They are windows into stories, information, and memories. Yet, not all images are born equal. Sometimes they lack the clarity and appeal needed to truly captivate us, making image enhancement a remarkable realm of technology that brings out the best in our visual experiences. Think about the moments when one struggles to make out details in a dimly lit photograph or needed to decipher critical information from a grainy surveillance video. This is where image enhancement techniques step in, like digital magicians, transforming these subpar visuals into something vibrant, crisp, and meaningful. They have the power to amplify contrast, sharpen edges, and eliminate distracting noise, turning ordinary images into extraordinary stories. Our journey in this paper is all about exploring these enchanting techniques, peeling back the curtain on their inner workings, and showing just how versatile and transformative they can be. From traditional filter-based methods to cutting-edge deep learning solutions, we get to reveal the secrets that breathe new life into images across various domains, including medicine, remote sensing, art, and more. But image enhancement isn't all technical wizardry, it's a field filled with ethical questions and considerations, especially when it comes to preserving privacy and ensuring that the power of enhancement is wielded responsibly. In an age where visuals are a universal language, understanding the magic of image enhancement is more vital than ever. With this research, a human-centred perspective on this dynamic world is aimed to be provided, one that enhances not just images but also our connection to the visual stories that surround us every day.
II. DIVERSE IMAGE ENHANCEMENT TECHNIQUES
A variety of image enhancement techniques are employed, encompassing both traditional and modern approaches. Traditional techniques include spatial domain methods such as histogram equalization and spatial filtering, while frequency domain methods like Fourier and wavelet transforms are also utilized. More recently, deep learning-based methods have gained prominence, leveraging convolutional neural networks (CNNs) and generative adversarial networks (GANs) to automatically learn and enhance image features. Each technique is applied based on specific image characteristics and enhancement goals, allowing for flexibility and adaptability in improving image quality.
A. CNN (Convolutional Neural Networks)
Image enhancement is a crucial field in image research and processing, aiming to artificially enhance photo quality and visual appearance. Many photos end up with high noise and grain due to poor sensor quality and control, or with low sharpness. Convolutional Neural Networks (CNN) rely on image structure and texture to serve as differentiating features. CNN-based techniques for image denoising, image super resolution, contrast enhancement, improving images in low light conditions are commonly encouraged. CNNs can be used to judge image aesthetics, improve various parts of the image such as contrast, colors, highlights, shadows without losing true to life accuracy, and then use Super Resolution to improve sharpness and detail.
Denoising is a major concern in image processing, as noise interferes with the picture and degrades the quality of the image. Before the use of neural networks, the BM3D algorithm was used. A deep CNN architecture consisting of a 30 layer Convolutional-Deconvolutional model and Symmetric Skip Connections (SSC) between the alternate Conv-DeConv layers showed better results than the BM3D algorithm. Low light image enhancement is another area where CNNs can be used. Using nearest neighbour classification to rebuild images, along with an HDR photography approach similar to Google RAISR's is one of the proposed approaches. (1)
B. Deep Fusion Network (DFN)
An end-to-end deep fusion-based approach is proposed to enhance the quality of images acquired in weak illumination environments. The proposed deep fusion network (DFN) employs a convolutional neural network (CNN) to generate confidence maps as spatial weighting factors for fusing images created by multiple base image enhancement techniques that complement each other in a content-dependent manner. Tests conducted on both synthetic and real weakly illuminated images demonstrate that the DFN approach delivers superior performance in terms of both subjective visual perception and objective quality assessment. High image quality is imperative in computer vision applications, such as object detection and image classification.
Existing methods can be categorized into three groups: histogram-based, Retinex-based, and learning-based. Histogram-based methods enhance image contrast and brightness by remapping pixel values in image intensity histograms. The Retinex model, describing an image as the product of illumination and reflectance, enhances degraded images by lightening the illumination component. Learning-based methods, exemplified by LLNet and CNN, utilize convolutional neural networks to estimate coarse illumination and smooth it using a guided filter. Experiments conducted on both synthetic and real weakly illuminated images reveal that DFN enhances image brightness and contrast while better preserving structural details compared to state-of-the-art image enhancement algorithms. DFN-enhanced images demonstrate superior performance in terms of both perceptual visual quality and objective image quality assessment. (2)
C. Histogram Equalization (HE)
A new method for enhancing micro-texture images using bi-histogram equalization based on arcsine distribution is present. The aim of this method is to address major issues associated with histogram equalization, such as over-enhancement and mean brightness shifting, which can result in an unnatural appearance. In this approach, each colour channel histogram is divided into two sub-histograms based on the mean threshold, and their cumulative distribution is replaced with arcsine distribution. Experimental results demonstrate that the proposed method outperforms other state-of-the-art methods.
Image enhancement, a critical task in image processing, is focused on enhancing image quality, especially when images suffer from factors like poor lighting, lens aperture, sensor sensitivity, and inexperienced camera operators. An effective contrast enhancement algorithm for micro-texture images should possess attributes such as simplicity in implementation, noise tolerance, brightness preservation, time efficiency, and robustness. The most widely used method in this context is Histogram Equalization (HE), which involves the creation of a nonlinear intensity function for images. However, it has a significant drawback due to its tendency to over-enhance images. Some refinements have been introduced, like Brightness Preserving Bi-Histogram Equalization (BBHE). (3)
D. General Methods and Mathematical Principles
There exists various general methods and mathematical principles used for detail enhancement in infrared image visualization. The split method, which adopts a guided filter to split the image into detail components and base components, has a side effect in compression due to increasing detail components' gray value. A simplified local processing measure based on image matting is designed to lower the influence. The gradient domain method revised a gradient gain factor function resulting from gray value overflow. The method cited solves energy function with high computational complexity and great occupation in memory.
A matting method can be proposed and tested to save more than 30% of time in solving system of linear equations and more than 70 times in matrix storage. Infrared image detail enhancement (DDE) algorithms are based on establishing a mapping system or covering the defect of the mapping system, which accommodates 14-bit large variation to 8-bit in an obvious and high contrast way. The goal is to obtain higher contrast and less detail loss. DDE approaches can be categorized into three subgroups: histogram projection, split method and gradient domain methods. In this context, the split method is complemented with histogram projection, presenting conflicts as a consequence of chasing detail and contrast enhancement simultaneously. The gradient domain method minimizes the energy function and solves a large system of linear equations. (4)
III. IMAGE ENHANCEMENT ACROSS DIVERSE IMAGE TYPES
The applicability of image enhancement techniques across various image types is explored in this section. The focus is on the versatility and adaptability of image enhancement methods, regardless of the image's source or characteristics. Image enhancement plays a critical role in improving image quality and revealing latent information, and its effectiveness is examined across a spectrum of image domains, including medical imaging, satellite imagery, artistic photography, and more. Understanding the broad range of scenarios in which image enhancement proves valuable contributes to its wider utilization and underscores its significance in enhancing visual content.
A. Hyperspectral Images (HSI)
A novel method for hyperspectral image filtering and enhancement is developed under the Partial Differential Equations framework. This method demonstrates notable noise filtering, details preservation, and enhancement capabilities. Its efficiency has been demonstrated both in terms of visual assessment and overall classification accuracy on several public hyperspectral image datasets when employed as a preprocessing step. High-resolution hyperspectral images (HSI) find applications in various processing systems, encompassing domains such as astronomy, agriculture, imaging, and geoscience. These images provide discriminative information about materials or vegetation based on their spectral signatures.
The denoising process for hyperspectral images is designed to eliminate noise in homogeneous regions while preserving edges and high-frequency details. Several denoising methods have been introduced to address HSI data by capitalizing on the correlation between the spatial and spectral dimensions. A model named the weighted hyperspectral total variation (TV) model (CSSWHTV) is proposed, which combines the spectral and spatial dimensions. Additionally, a pseudo-3D technique, featuring a coupling term between different channels, allows for the integration of inter-channel information into the filtering process, accounting for distinct spectral properties and their correlations. (5)
B. Near Infrared Imaging (NIR)
The aim of this study is to test orthogonal projection as a spectral pre-treatment method for reducing the interference of polystyrene signal in Near Infrared (NIR) imaging of agar on plastic petri dishes. Polystyrene plastic petri dishes are commonly used as sample holders in microbiological studies due to their affordability, disposability, and standardized appearance, but their use is generally restricted in NIR spectral imaging due to the spectral signature of the dishes. A combination of Savitzky Golay smoothing, SNV, and orthogonal projection is employed to mitigate the spectral differences between agar samples presented on polystyrene and glass dishes. This approach diminishes the impact of polystyrene dishes on samples with varying agar thicknesses. Subsequent research will assess the viability of this approach as a method for aiding the identification of bacteria on polystyrene petri dishes through NIR hyperspectral imaging. The study centres on the effect of petri dish material and cover on the spectral profile of agar on plastic and glass petri dishes, both with and without a glass lid. Additionally, the influence of plastic petri dishes is examined in the context of different agar thicknesses. The spectral characterization of sample presentation and the impact of spectral pre-treatments on the spectral features of diverse sample presentations can serve as a valuable resource for researchers dealing with samples presented on agar petri dishes. (6)
Image enhancement techniques have far-reaching applications that extend beyond the confines of image processing. In a world increasingly reliant on visual information, the ability to enhance image quality and clarity has become indispensable. These techniques find utilization in diverse domains, such as medical diagnostics, surveillance, remote sensing, photography, art, and more. The applications of image enhancement are as varied as the domains they serve, making them integral to enhancing visual content for both technical analysis and aesthetic appreciation. In this context, this paper explores the multifaceted applications of image enhancement techniques and their transformative impact across different fields.
A. Identify Bacterial Colonies
The objective is the detection and quantification of bacteria colonies within the domain of medical microbiology without encountering any limitations. An image-processing-based bacteria colony counter was devised in MATLAB, facilitating an investigation into three categories of bacteria responsible for hospital-acquired infections: Escherichia coli, Pseudomonas aeruginosa, and Enterococcus faecalis.
The Circular Hough Transform (CHT) in MATLAB was harnessed for the detection and enumeration of bacteria colonies, accompanied by the development of a Graphical User Interface (GUI) for enhanced practicality.
Microbials, the smallest living organisms, are subjects of scientific exploration regarding their relationships with other living entities. Microbiology has seen significant evolution over time, with essential treatment modalities like vaccines and antibiotics assuming paramount importance. Six prevalent diagnostic methods are employed in the identification of bacterial infections, including morphological examination, biochemical reactions, antibiotic susceptibility tests, antibody detection, assays for antigens and nucleic acids, and isolation, culture, and identification. This study adopts the isolation, culture, and identification approach to scrutinize bacteria colonies on agar surfaces. Following manual examination of petri dishes, digital image processing is applied to enhance visual quality and execute mathematical operations on the images. The digital image processing encompasses three primary phases: image acquisition, pre-processing, and description. Four categories of image processing algorithms are enlisted: histograms, filters, convolution algorithms, and mathematical operations. (7)
B. Automatic Counting for Nuclei Cells
An automatic counting method for nuclei cells in histological images is implemented, utilizing efficient image processing techniques. The method relies on the utilization of image thresholding, morphological image processing operations, and the connected component algorithm. High accuracy, up to 89.5%, was observed in experimental results when compared to other previous research work. The approach combines machine learning techniques with various image processing methods such as the watershed transform, distance transform, k-means, adaptive contour, Bayesian, and support vector machine (SVM).
The manual counting of nuclei cells in histological images is widely recognized as a laborious and error-prone task. The automation of this process is essential for enhancing the analysis of histological images. Existing systems and methodologies primarily rely on color or grayscale images, which often yield inaccurate outcomes and exhibit multiple limitations. The proposed approach was empirically assessed using a dataset of 37 images from a publicly available collection of 100 histological images. The experimental results clearly highlight the efficacy of the proposed method for the automated counting of nuclei cells in histological images. (8)
C. Classify Bacterial Images
The goal is to develop a system capable of automatically identifying and classifying bacteria images in standard resolution using image classification and deep learning techniques. An implementation approach for a bacteria recognition system is proposed, utilizing Python programming and the Keras API in conjunction with the TensorFlow Machine Learning framework. The results of the implementation confirm the system's ability to recognize the genus of bacteria in microscope images. The experimental results compare the accuracy of deep learning methods in the context of bacteria recognition with standard resolution images.
This method has the potential to be applied to both high-resolution and standard resolution datasets for predicting bacteria types. Two species of bacteria, namely Staphylococcus aureus TISTR 746 and Lactobacillus Delbrueckii TISTR 1339, were used, employing deep learning techniques through Python programming. The LeNet Convolutional Neural Network (CNN) architecture was utilized for training and testing datasets comprising over 400 sample images. The objective is to develop a system that can promptly and automatically identify and classify bacteria images in standard resolution. The proposed approach can be extended to include high-resolution and standard resolution datasets for predicting bacteria types. (9)
D. Rapid Identification of Mycobacteria
A straightforward method for the swift identification of Mycobacteria species using MALDI-TOF MS (Matrix-Assisted Laser Desorption/Ionization-Time of Flight Mass spectrometry) with the Bruker MALDI-TOF Biotyper system is present. A multicentre, prospective, and single-blind investigation was conducted in three European Hospitals, comprising two Spanish hospitals and one UK hospital, spanning from May to August 2018.
The BD BACTEC MGIT liquid culture system was uniformly employed in all three centres for the cultivation of Mycobacteria. Upon the generation of a positive signal, the test tubes were removed from the analyzer and subjected to subculture on blood agar plates for subsequent MALDI-TOF analysis. The plates were incubated aerobically for a period ranging from 1 to 7 days at 37°C and were examined daily. As soon as visible growth was detected, it was transferred onto a steel target plate, where it was overlaid with neat formic acid and 1 μl HCCA matrix (alpha hydroxyl 4 cinnamic acid), and subsequently analyzed using a Bruker Biotyper MALDI-TOF instrument. The results obtained through MALDI-TOF were compared with the reference methodologies employed for identification in the different centres.
In total, in 142 out of 167 cases (85%), the identifications achieved were found to be in agreement; specifically, all Mycobacterium tuberculosis (MTB) isolates (43 out of 43, 100%), 57 out of 76 (75%) of the rapidly growing nontuberculous mycobacteria (NTM), and 42 out of 48 (85%) of the slow-growing NTM that were tested were correctly identified. This novel, uncomplicated, cost-effective, and rapid method for the isolation and identification of Mycobacterium species is now routinely utilized across all three centres. The prompt identification of mycobacterial infections and the detection of antimicrobial resistance continue to represent significant challenges for the microbiology laboratory. The definitive identification turnaround time typically relies on external reference laboratory services, which can result in substantial delays. (10)
 A Comprehensive Review of Convolutional Neural Network based Image Enhancement Techniques by Katyani Singh, Avi Seth, Harshit Singh Sandhu and Prof. Krishna Samdani.  Enhancement of Weakly Illuminated Images by Deep Fusion Networks by Yu Cheng, Jia Yan and Zhou Wang.  Enhancement of Micro-texture Images Using Bi-histogram Equalization Based on Arcsine Distribution by Suresha M, Raghukumar DS and Kuppa S  Analysis and Improvement of General Methods in Infrared Image Detail Enhancement by Yitian Li, Chengguang Ma, Zhizhe Liu and Ziqi Liu.  Hyperspectral Image Enhancement using Diffusion and Shock Filtering Techniques by Andreia Valentina Miclea, Monica Borda, Romulus Terebes and Serban Meza.  Orthogonal Projection as A Spectral Pre-Treatment Method to Reduce The Interference Of Polystyrene Signal In NIR Imaging Of Agar On Petri-Dishes by Ana Herrero-Langreo, Nathalie Gorretta, Agnes Beghin, Yu-Mei, Mariateresa Ferone, Aoife Gowen, Amalia Scannell.  Counting Bacteria Colonies Based on Image Processing Methods by Busra KIS, Mazlum UNA Y and Gizem Dilara, Utku Kursat ERCAN and Aydin AKAN.  An Automatic Nuclei Cells Counting Approach Using Effective Image Processing Methods by Mogeeb A. A. Mosleh and Abdul Aziz AL-Yamni, Abdu Gumaei.  Bacteria Classification using Image Processing and Deep learning by Treesukon Treebupachatsakul and Suvit Poomri?gul.  An improved simple method for the identification of Mycobacteria by MALDI-TOF MS (Matrix-Assisted Laser Desorption- Ionization mass spectrometry) by AdelaAlcolea-Medina, M.T. Cabezas Fernandez, N. Montiel, M. P. LuzónGarcía, C. Delamo Sevilla, Nathan North, M. J. Martinez Lirola and MarkWilks.
Copyright © 2023 Aishwarya Jagadish. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.