52039 research outputs found
Sort by
PixLabelCV - Labeling images for semantic segmentation fast, pixel-precise and offline
No full textImage annotation, also called labeling is a necessary task for any supervised learning approach to obtain ground
truth data for model training. This article offers a comprehensive survey of contemporary image annotation tools,
grouping freely accessible ones based on their service range, speed, and data privacy assurances.
In our exploration for tools capable of executing pixel-precise semantic labeling, we identified a shortage of swift,
free image annotation tools that don’t require users to upload their data to third-party servers. Therefore, we
introduce "PixLabelCV" - a lightweight, fast, offline, and standalone annotation tool primarily developed to aid
human annotators in achieving pixel-perfect labels promptly. Uniquely crafted to be freely available (open source)
and non-server-based, it ensures enhanced privacy and efficiency. Hence, it is aimed to serve as an ideal tool to
facilitate labeling data for smaller labs and businesses.
At its core, PixLabelCV fuses conventional labeling techniques such as delineating objects with rectangles or
polygons with multiple computer vision algorithms. Spanning basic thresholding in RGB or HSV color space to
more intricate procedures like flood fill or watershed the tool instantaneously computes and exhibits the resulting
segmentations. Annotators can swiftly add these segments to a class label or refine them by adjusting parameters
or markers before a quick repetition. To further augment the user experience, additional functionalities like mor phological closing are incorporated, facilitating an intuitive labeling process. Another standout feature is its ability
to uniquely assign pixels to singular classes, eliminating any potential overlap-induced ambiguities
Combining bidirectional path tracing, DDGI, and ReSTIR to improve real-time rendering quality
No full textThis paper introduces a new algorithm for real-time rendering that builds upon bidirectional path tracing, reservoir based spatio-temporal importance resampling, and dynamic diffuse global illumination. The combination of these
algorithms produces an image with reduced noise compared to real-time run algorithms like path tracing, while
retaining details like caustics. The resulting darkening, which is discussed in greater detail in this paper, is also
reduced due to the usage of importance resampling of points on light-emitting surfaces. While the standard al gorithms, such as bidirectional path tracing, cannot be run in real-time with satisfactory quality, a set of novel
approaches have emerged to fill this gap. These algorithms are capable of running in real-time, although they
do suffer from certain limitations. This paper describes the combination of bidirectional path tracing, DDGI and
ReSTIR. This rectifies the drawbacks of missing indirect reflection, darkening and missing caustics of these al gorithms. Ultimately, all the results of these algorithms are compared by verifying real-time rendering time and
comparing quality to reference images. The quality is evaluated by using comparions for darkening and the sim ilarity of the real-time rendered result to an offline path-traced result. The results of this paper demonstrate that
the algorithm presented improves upon previous algorithms in terms of quality, while still maintaining real-time
rendering constraints
Deep learning-based classification of breast tumors using selected subregions of lesions in sonograms
No full textBreast cancer, a prevalent disease among women, demands early detection for better clinical outcomes. While
mammography is widely used for breast cancer screening, its limitation in e.g., dense breast tissue necessitates
additional diagnostic tools. Ultrasound breast imaging provides valuable tumor information (features) which are
used for standardized reporting, aiding in the screening process and precise biopsy targeting. Previous studies have
demonstrated that the classification of regions of interest (ROIs), including only the lesion, outperforms whole
image classification. Therefore, our objective is to identify essential lesion features within such ROIs, which are
sufficient for accurate tumor classification, enhancing the robustness of diagnostic image acquisition. For our
experiments, we employ convolutional neural networks (CNNs) to first segment suspicious lesions’ ROIs. In a
second step, we generate different ROI subregions: top/bottom half, horizontal subslices and ROIs with cropped out center areas. Subsequently these ROI subregions are classified into benign vs. malignant lesions with a second
CNN. Our results indicate that outermost ROI subslices perform better than inner ones, likely due to increased
contour visibility. Removing the inner 66% of the ROI did not significantly impact classification outcomes (p =
0.35). Classifying half ROIs did not negatively impact accuracy compared to whole ROIs, with bottom ROI
performing slightly better than top ROI, despite significantly lower image contrast in that region. Therefore, even
visually less favorable images can be reliably analyzed when the lesion’s contour is depicted. In conclusion, our
study underscores the importance of understanding tumor features in ultrasound imaging, supporting enhanced
diagnostic approaches to improve breast cancer detection and management
From Sources to Solutions: Enhancing Object Detection Models through Synthetic Data
No full textObject detection, a fundamental task in computer vision, plays a crucial role in various applications such as au tonomous driving, surveillance, and robotics. However, training models for this task require vast amounts of
high-quality data, often involving labor-intensive manual labeling. Synthetic data, a promising alternative, re mains an active area of research. This paper presents a comprehensive exploration of different object sources
for the use of synthetic data in enhancing object detection models. We investigate various synthetic data gen eration techniques to implant objects into a scene, with a focus on enhancing training data diversity. These ob jects are either gathered from the training dataset itself using SegmentAnything as a new supervised self aug mentation technique or imported from external sources, including a photobox with a rotating table and web
scraping of online shops. Moreover, our study delves into the development of a placement logic that gradu ally evolves from placing objects randomly to placing objects in physically correct orientations to mimic the
real world data. We investigate the use of different blending techniques. The outcome of our study demon strates that synthetic images, when integrated with an existing real training set, substantially improve the ob ject recognition accuracy of the model without compromising inference time. Our code can be found at
https://github.com/EduardBartolovic/synthetic-data-generatio
ECNNXAI: Ensembled CNNs with eXplainable Artificial Intelligence for Colon Histopathology Image Classification
No full textColon cancer is ranked as the third most commonly diagnosed cancer and second for causing the most cancer
related deaths. Histopathology is a crucial diagnostic tool for cancer since it enables the microscopic analysis of
tissue samples to pinpoint abnormal cells, to identify the stage of the cancer and its kind. There is a significant need
for precise detection and diagnosis from histopathology images. This research proposes a stacking ensemble model
called Ensembled Convolutional Neural Networks with eXplainable Artificial Intelligence (ECNNXAI) for mul ticlass colon histopathology image classification. Our ensemble model consists of three pre-trained convolutional
neural networks (XceptionNet, DenseNet-121 and InceptionNetV3) as base classifiers and the logistic regression
as the meta classifier. Gradient-weighted Class Activation Mapping (Grad-CAM) visualization technique is used
to interpret and understand the regions focused by the base classifiers to arrive at the final predictions. SHapley
Additive exPlanations (SHAP) is used for understanding the predictions made by the ECNNXAI. The proposed
model achieves the best overall performance with accuracy of 72.83%, precision of 77.78%, recall of 66.52% and
F1 score of 71.71% on the Chaoyang dataset
Polychromatism of all light waves and a new approach to the interpretation of fluorescence mechanisms; Idea of the fluorescent color monitor
No full textIn scientific literature it is considered that only white light is polychromatic and that it is composed of countless
monochromatic waves. This research on light vision mechanisms in biosystems and on the mechanisms of
formation of deficits in color discrimination reveals that not only white light is polychromatic but all light waves
are. This hypothesis brings numerous consequences, two of which we want to address in this work:
- what does the measurement of wavelength mean when the radiation is polychromatic;
- the presence of wavelengths shorter or longer than the incident one must not mean Stokes and anti-Stokes
emissions but can be interpreted as the selective absorption/reflection mechanism of a compound radiation.
Based on these considerations, the mechanisms of fluorescence are interpreted as:
1. Pseudo-fluorescence formed by the selective absorption/reflection of consecutive monochromatic
components of polychromatic blue/violet light – spectral colors production;
2. Fluorescence formed by the selective reflection of non-consecutive monochromatic components of
polychromatic blue/violet light – fluorescent color production.
This new approach to the interpretation of fluorescence mechanisms is the theoretical starting point for project a
fluorescent colors monitor
A Comparative Study of Convex Combination and Inner Ordinate Methods For Scattered Data Interpolation Using Quartic Triangular Patch
No full textIn this study, we perform a comparative evaluation and assessment for the scattered data interpolation using a
quartic polynomial triangular patch with ten control points on a triangular domain. The comparison is made
using two different convex combinations and inner ordinates methods, i.e., cross derivative and cubic precision.
Statistical Goodness-fit measurements used are maximum error, coefficient of determination (R
2
), CPU time (in
seconds), and contour plot. From the result, the cubic precision method with linear convex combination methods
gave better results with smaller CPU times and higher R
2 value. All numerical and graphical results are presented
using MATLAB programmin
Analysis of natural lighting conditions for the digitization of artwork in an art gallery interior
No full textThe paper discusses the analysis of natural lighting conditions for digitizing art. The emphasis is on a realistic
3D digital reproduction of a work of art in natural lighting conditions in the interior of an art gallery. The art
object is scanned and digitized in two natural lighting conditions. The photogrammetry method was used for a
realistic 3D reconstruction of the artwork. This experiment aims to analyze the influence of lighting conditions
on the quality of 3D reproduction of an art object concerning image processing and color reproduction. In this
study, no accessories were used to increase the quality of the captured image, such as reflective and diffusive
plates or lights to illuminate the art object. The art object was scanned and digitized in two natural lighting
conditions. This study aims to analyze the influence of actual lighting conditions on the quality of a realistic
digital 3D reproduction of a work of art
The color recognition methods for the active markers in the motion capture system, using various techniques, including ML
No full textThe article focuses on a method for reliably identify moving colored artificial markers in real-time. The
marker was used to determine the 3D position in the space of the user(s).
The goal was to ensure that points were found and identified predictably and reliably by many cameras
simultaneously, which, with appropriate calibration, merging, and processing of the data, could provide
reliable information about the current 3D position of a given point in real-time. This information was
crucial to other components of the broader vision system (VR platform).
The problems encountered and the remedial methods discussed in the presentation concern several aspects
that we encountered during research, such as changes in lighting conditions, the quality (and stability) of
the generated light and color, the dependence of color recognition on the distance of the light source from
the camera matrix, aspects of light reflections, and many others. During our research, we analyzed various
RGB/RGBW LED light sources from different manufacturers, which are characterized by different light
generation characteristics. We also used a light diffuser. Using different sets of cameras and lighting
conditions, we conducted several studies and experiments.
During the research, we managed to find basic colors for our marker-tracking visual system that met the
goals. We have proposed an algorithm to deal with the problem and demonstrate the reliability of the visual
layout with the algorithm. During our research, we used both conventional and alternative techniques
related to ML
Impact of the State-of-the-Art Methods on Camera Trap Image Classification
No full textThe work has been supported by the grant of the University of West Bohemia, project
No. SGS-2022-017. Computational resources were provided by the e-INFRA CZ project
(ID:90140), supported by the Ministry of Education, Youth and Sports of the Czechia