The latest World Health Organization’s (WHO) study on 2018 is showing that about 1.5 million people died and around 10 million people are infected with tuberculosis (TB) each year. Moreover, more than 4,000 people die every day from TB. A number of those deaths could have been stopped if the disease was identified sooner. In the recent literature, important work can be found on automating the diagnosis by applying techniques of deep learning (DL) to the medical images. While DL has yielded promising results in many areas, comprehensive TB diagnostic studies remain limited. DL requires a large number of high-quality training samples to yield better performance. Due to the low contrast of TB chest x-ray (CXR) images, they are often is in poor quality. This work assesses the effect of image enhancement on performance of DL technique to address this problem. The employed image enhancement algorithm was able to highlight the overall or local characteristics of the images, including some interesting features. Specifically, three image enhancement algorithms called Unsharp Masking (UM), High-Frequency Emphasis Filtering (HEF) and Contrast Limited Adaptive Histogram Equalization (CLAHE), were evaluated. The enhanced image samples were then fed to the pre-trained ResNet and EfficientNet models for transfer learning. In a TB image dataset, we achieved 89.92% and 94.8% of classification accuracy and AUC (Area Under Curve) scores, respectively. All the results are obtained using Shenzhen dataset, which are available in the public domain.

/pdf/image-enhancement-for-tuberculosis-detection-using-deep-3q3l84tefi.pdf

Image Enhancement for Tuberculosis Detection Using Deep Learning

Feature matching is a challenging computer vision task that involves finding correspondences between two images of a 3D scene. In this paper we consider the dense approach instead of the more common sparse paradigm, thus striving to find all correspondences. Perhaps counter-intuitively, dense methods have previously shown inferior performance to their sparse and semi-sparse counterparts for estimation of two-view geometry. This changes with our novel dense method, which outperforms both dense and sparse methods on geometry estimation. The novelty is threefold: First, we propose a kernel regression global matcher. Secondly, we propose warp refinement through stacked feature maps and depthwise convolution kernels. Thirdly, we propose learning dense confidence through consistent depth and a balanced sampling approach for dense confidence maps. Through extensive experiments we confirm that our proposed dense method, \textbf{D}ense \textbf{K}ernelized Feature \textbf{M}atching, sets a new state-of-the-art on multiple geometry estimation benchmarks. In particular, we achieve an improvement on MegaDepth-1500 of +4.9 and +8.9 AUC$@5^{\circ}$ compared to the best previous sparse method and dense method respectively. Our code is provided at https://github.com/Parskatt/dkm

pdf/dkm-dense-kernelized-feature-matching-for-geometry-ndmlw1lv.pdf

DKM: Dense Kernelized Feature Matching for Geometry Estimation

Image enhancement is an essential technique used in many imaging applications. The main motivation of image enhancement is processing an image to be more suitable for specific utilization. Various image enhancement methods have been proposed in the existing literature. On the one hand, a variety of enhancement techniques make the classification more complex. On the other hand, the lack of image quality metrics specifically designed for enhanced images and the number of enhanced image quality databases as well as subjective data (ground truth) is limited, making the evaluation process difficult. This survey aims to guide the image enhancement research community to fill the above-mentioned gap and develop better and robust enhancement techniques. This survey 1) summarizes previous surveys of image enhancement techniques and existing classifications; 2) classifies the amount of techniques and analyzes their components, including how the performance has been evaluated; 3) surveys and recommends image quality metrics and databases that were designed for enhanced images; 4) discusses the challenges for image enhancement and its evaluation issues; and 5) proposes perspectives for the development of future image enhancement algorithms. This survey mainly focuses on image enhancement techniques that improve the perceived quality of natural images. 

Survey of natural image enhancement techniques: Classification, evaluation, challenges, and perspectives

Dense feature matching is an important computer vision task that involves estimating all correspondences between two images of a 3D scene. In this paper, we revisit robust losses for matching from a Markov chain perspective, yielding theoretical insights and large gains in performance. We begin by constructing a unifying formulation of matching as a Markov chain, based on which we identify two key stages which we argue should be decoupled for matching. The first is the coarse stage, where the estimated result needs to be globally consistent. The second is the refinement stage, where the model needs precise localization capabilities. Inspired by the insight that these stages concern distinct issues, we propose a coarse matcher following the regression-by-classification paradigm that provides excellent globally consistent, albeit not exactly localized, matches. This is followed by a local feature refinement stage using well-motivated robust regression losses, yielding extremely precise matches. Our proposed approach, which we call RoMa, achieves significant improvements compared to the state-of-the-art. Code is available at https://github.com/Parskatt/RoMa

RoMa: Revisiting Robust Losses for Dense Feature Matching

An effective contrast enhancement method should not only improve the perceptual quality of an image but should also avoid adding any artifacts or affecting naturalness of images. This makes Contrast Enhancement Evaluation (CEE) a challenging task in the sense that both the improvement in image quality and unwanted side-effects need to be checked for. Currently, there is no single CEE metric that works well for all kinds of enhancement criteria. In this paper, we propose a new Multi-Criteria CEE (MCCEE) measure which combines different metrics effectively to give a single quality score. In order to fully exploit the potential of these metrics, we have further proposed to apply them on the decomposed image using wavelet transform. This new metric has been tested on two natural image contrast enhancement databases as well as on medical Computed Tomography (CT) images. The results show a substantial improvement as compared to the existing evaluation metrics. The code for the metric is available at: https://github.com/zakopz/MCCEE-Contrast-Enhancement-Metric

A Multi-Criteria Contrast Enhancement Evaluation Measure using Wavelet Decomposition

We tackle the essential task of finding dense visual correspondences between a pair of images. This is a challenging problem due to various factors such as poor texture, repetitive patterns, illumination variation, and motion blur in practical scenarios. In contrast to methods that use dense correspondence ground-truths as direct supervision for local feature matching training, we train 3DG-STFM: a multi-modal matching model (Teacher) to enforce the depth consistency under 3D dense correspondence supervision and transfer the knowledge to 2D unimodal matching model (Student). Both teacher and student models consist of two transformer-based matching modules that obtain dense correspondences in a coarse-to-fine manner. The teacher model guides the student model to learn RGB-induced depth information for the matching purpose on both coarse and fine branches. We also evaluate 3DG-STFM on a model compression task. To the best of our knowledge, 3DG-STFM is the first student-teacher learning method for the local feature matching task. The experiments show that our method outperforms state-of-the-art methods on indoor and outdoor camera pose estimations, and homography estimation problems. Code is available at: https://github.com/Ryan-prime/3DG-STFM.

3DG-STFM: 3D Geometric Guided Student-Teacher Feature Matching

The quality of images or videos that suffer from exposure distortions can be enhanced using histogram equalization or retinex methods. However, the relationship between the visual quality and the degree of enhancement is an inverted U-shaped function with a peak point, and many existing methods have parameters that have no clear relationship with image quality. We introduce a controllable illumination enhancement system, where the degree of enhancement can be adjusted using a single parameter. We then propose an over-enhancement measure, Lightness Order Measure (LOM), which quantifies the unnaturalness based on a local inversion of lightness order. We explore the relationship between the peak point and LOM in a subjective test. The results indicate that LOM reduces content dependency compared to existing methods. Our subjective test also evaluates the image quality of our enhancement, and demonstrates the effectiveness of our method.

Controllable Image Illumination Enhancement with an Over-Enhancement Measure

First-person videos (FPVs) captured by wearable cameras often contain heavy distortions, including motion blur, rolling shutter artifacts and rotation. Existing image and video quality estimators are inefficient for this type of video. We develop a method specifically to measure the distortions present in FPVs, without using a high quality reference video. Our local visual information (LVI) algorithm measures motion blur, and we combine homography estimation with line angle histogram to measure rolling shutter artifacts and rotation. Our experiments demonstrate that captured FPVs have dramatically different distortions compared to traditional source videos. We also show that LVI is responsive to motion blur, but insensitive to rotation and shear.

Characterizing distortions in first-person videos

Image quality assessment in first-person videos

First-person videos (FPVs) recorded by wearable cameras have different characteristics compared to mobile videos. Video frames in FPVs are subject to blur, rotation, shear and fisheye distortions. We design a subjective test that uses actual captured images with real distortions, synthetic distortions or a combination of both. Results indicate shear is less sensitive to content than rotation. For fisheye, personal preference and content dependence affect the subjective results. The performance of 7 noreference (NR) quality estimators (QEs) and our QE, local visual information (LVI) [1], are evaluated based on subjective results. We propose two mapping functions for rotation and shear that improve the ability of LVI and 4 NR QEs to accurately predict the subjective scores. Introduction As wearable cameras gain increasingly popularity, the socalled first-person videos (FPVs) grow to be an important video source over networks [2]. FPVs have different attributes compared to mobile videos. When recording mobile videos, people can see the ongoing scene in the screen, so they can be aware of how to fit the scene into the camera shot and intentionally control the video to be less shaky. However, FPVs are passive videos, and often lack real-time feedback about the field of view. As a result, the wearer rarely adjusts the camera to be stable or horizontal. Recently, the increased volume of uploading, transmission and storage of FPVs has given rise to a variety of applications. Human Activities in FPVs are classified in [3]. The detection of “snap points” was proposed to extract frames that are worthy to be captured and stored [4]. The ego-sampling method was developed to sample useful frames for faster video browsing [5]. Because of the capture process of FPVs, quality degradations are not limited to transmission or post-processing, and the resulting distortions in frames have not been subjectively evaluated. Motion blur and geometric distortions are two major distortions in FPVs [1]. Motion blur mainly arises from fast motion of the camera. When the camera keeps changing its positions in the scan time of one frame, the captured scene is blurred. Geometric distortions can be classified into 3 categories: rotation, shear and fisheye. Rotation results from head or body rotation. Wearers regularly move their bodies and shake their heads, and rarely are aware whether or not the camera is kept horizontal while recording videos. Shear is caused by camera panning. When the camera changes its positions in the scan time of one frame, the top rows of the frame are not vertically align with the bottom rows. For example, architecture in a sheared image is visually skewed. Fisheye images are captured by wearable cameras with ultra wide-angle lens (i.e. Gopro, Looxie Camera, Mobius). Instead of capturing a rectilinear image, the content appears to be convex. Fisheye is one of the lens distortions, called barrel transformation. The transformation warps the image to be bent; the magnification decreases from center to margins [6]. Most existing subjective tests [7, 8, 9] explored quality degradations starting from an undistorted image, which is considered to be the reference image. By synthetically adding distortions (i.e. Gaussian blur, JPEG, JPEG 2000, noise) to the reference image with different degradation levels, a series of distorted images of the same content but different severity are created. To design a subjective test for FPVs, however, applying each type of distortions separately fails to consider two issues: 1. Degradations in actual images captured by a camera may be subtly different than those created synthetically using a model [10]. In our test, images with real, synthetic or real plus synthetic distortions are evaluated. The real images are extracted from frames in FPVs. 2. Many blurry images are subject to geometric distortions (i.e. rotation, shear or fisheye) simultaneously in FPVs. The question is what is the visual impact on the overall quality when one image has multiple distortions. Multiply-distorted images have been evaluated in the subjective test, but only for blur, JPEG and noise [11]. In addition, these distortions, like blur, JPEG or noise, all have pixel-to-pixel correspondence, whereas geometric distortions do not. When constructing images for the subjective test for different amounts of geometric distortions, we keep the same content in the image center to maintain a consistent focus of attention. Actual images captured by a wearable camera have no reference image, so full-reference (FR) quality estimators (QEs) [12, 13] and reduced-reference (RR) QEs [14, 15] cannot be used to measure their quality, but no-reference (NR) QEs [16, 17, 18] are commonly content dependent and show very different behaviors for different distortions applied to the same image. Using our subjective test, we evaluate the behaviors of different NR QEs for actual captured images and multiply-distorted images. We design a subjective test to evaluate motion blur and geometric distortions in FPVs. Section 2 compares existing QEs with our QE, local visual information (LVI) [1]. The strategy of the quality estimation in LVI cannot be categorized into FR, RR or NR methods. In section 3, our subjective test evaluates actual captured images with real distortions, synthetic distortions or a combination of both using the paired comparison method. The types of distortions include motion blur, rotation, shear and fisheye. In section 4, we proposed two mapping functions for rotation and shear to compute the overall quality of images with blur and geometric distortions. Personal preferences and content dependence in fisheye are discussed. Comparison of LVI with existing QEs FR QEs and RR QEs often use a reference image to estimate the quality of a distorted image with pixel-to-pixel correspondence. They cannot be used for any image for which there is no corresponding reference. NR QEs uses only a single image without any reference to compute a quality score. In [1], we introduced a QE, LVI, that does not fit in the typical categorization of QEs into FR, RR and NR [19]. LVI estimates the quality of a test image based on a pseudo-reference image. Unlike in FR QEs, the pseudo-reference image and the test image are not necessarily pixel-aligned but obey the constraint that they share overlapping content. For example, the two images could be captured by a camera in the same scene but different camera angles. The geometric relationship between the two images are usually modeled by a homography. Since the quality of the pseudo-reference image is unknown, LVI is a relative quality estimation instead of an absolute quality score as computed by FR or NR QEs. LVI primarily measures the amount of relative blur of the test image based on the pseudoreference. When the test image is more blurred than the pseudoreference image, its LVI score would be in the range from 0 to 1. Otherwise, its LVI score would be larger than 1, which indicates the test image is sharper. If the two images have too few matching feature points, we then set the LVI score of the test image to be 0. LVI is insensitive to rotation and shear [1]. To make LVI be able to measure geometric distortions, we need to include rotation and shear into its quality estimation. We design a subjective test to evaluate multiply-distorted (blur with geometric distortions) images. The subjective results are used to develop the rotation and the shear mapping functions for LVI. The mapping functions take the rotation or shear and the LVI score as inputs to compute the overall quality score of an image with simultaneous blur and geometric distortions. Subjective Test Our test employs the paired comparison method for still images containing both actual and synthetic distortions that are typical of images extracted from FPVs. Each pair of test images are simultaneously displayed on two monitors side by side. Viewing distance is kept to be around 3 times the height of test images. 9 videos have been recorded by a Pivothead camera (frame rate: 30fps, resolution: 1920×1080) including “billiards”, “eating”, “flight”, “bell tower”,“winter Hovde Hall”, “parking lot”, “autumn Hovde Hall”, “apartment building” and “parking garage”. 4 distortions including motion blur, rotation, shear and fisheye, are evaluated. Test images are either real frames from the 9 videos or created by adding synthetic distortions to selected frames. Because synthetic shear and fisheye change the image size compared to the original image, all test images are cropped to be 1600×900 to remove marginal regions with little content. Test Method In a paired comparison subjective test, the subject needs to indicate his or her preference among the two images according to their visual quality. 50 subjects including 43 males and 7 females participated in the test. All pairs of images are displayed in random order. To improve the efficiency of paired comparison, we use the “square design” in [20]. Given that we have n stimulus, a full comparison needs 0.5n(n− 1) pairs. By using the “square (a) (b) (c)

Subjective evaluation of distortions in first-person videos.

Handle everyday research tasks with reliable, citation-backed results

Your personal Research Agent to handle research tasks with citation-backed results

Popular Tasks used by Researchers

How can I help with your research?

Meet SciSpace

Get more enhanced response by uploading the PDFs you want me to reference.

No relevant PDFs in your library

SciSpace is the AI research assistant for academics. Run systematic literature reviews on 280M+ papers, and write papers with cited sources. Trusted by 1M+ students, PhDs & researchers.

SciSpace | AI for Research

Analyze PDFs

Code & Manuscripts

Funding & Grants

Literature & Patents

Medical & Clinical Data

Systematic Review

Visualize & Present

Web & Data

Build a Google Scholar-like website for your research.

Build a website

Create charts and images for your research

Create a Chart

Write a paper for submission to a journal

Draft a manuscript

Patent Search

Design eye-catching scientific posters in minutes.

Scientific Poster Generation

Systematic Literature Review

One task is running at the moment. Your messages will be shown right after.

Drag and drop or click here to browse

Loved by <highlight>1 million+</highlight> researchers

Extract a list of specific topics and their sources from unstructured text

Topics

Compare and analyze relevant papers that matches with your search

Papers

Get insights from PDFs and bookmarked papers from your library

My library

Recent searches

Try searching for:

Catch AI-generated content in scholarly and non-scholarly content

{ai} Detector

Ai Writer

Get PDF Summaries, highlighted text explanations 

Chat with PDF

Effortlessly create in-text citations and bibliographies in APA and 2,500 other formats

Citation generator

Get explanations, summaries, and answers on academic papers

Ease up your research workflow with {scispace}'s cohort of exciting AI tools

Elevate your academic writing skills and convey your ideas the way you want

Paraphraser

Explore our range of reading and writing tools

Your file is being prepared and should be ready in a few minutes. If it's a large file, it might take a bit longer. You can close this window, and we'll email you the file when it's done.

You have reached a maximum limit of <strong>{limit}</strong> columns in the table. Remove at least <strong>1</strong> column to add or create another one.

Chen Bai

Author Tools

Chat about Author