We consider the problem of reconstructing binary images from their row and column sums with prescribed number of strips in each row and column. In a previous paper we compared an exact deterministic and an approximate stochastic method (Simulated Annealing – SA) to solve the problem. We found that the latter one is much more suitable for practical purposes. Since SA is sensitive to the choice of the initial state, in this paper we present different strategies for choosing a starting image, and thus we develop variants of the SA method for strip constrained binary tomography. We evaluate the different approaches on images with varying densities of object pixels.
We associate a closure operator with every n-ary relation ( n>1 an integer) on a given set. We focus on certain n-ary relations on the digital line ℤ and study the closure operators on the digital plane ℤ^2 that are associated with special products of pairs of the relations. These closure operators, which include the Khalimsky topology, are shown to provide well behaved connectedness, so that they may be used as background structures on the digital plane for the study of digital images.
The occlusal splint is one of the methods of treatment of discrepancies between the centric relation and maximal intercuspation (CR/MI), and other temporomandibular joint (TMJ) disorders. It is also a method of reducing the effects of bruxism. Designing an occlusal splint for a given relation between the maxilla and the mandible involves: creating partial surfaces, integrating them, and producing the splint on a 3D printer. The paper presents and compares some techniques used to design splint surfaces under a required therapeutic maxilla-mandible relation.
We used model-free methods to explore the brain’s functional properties adopting a partitioning procedure based on cross-clustering. We selected Fuzzy C-Means (FCM) and Neural Gas (NG) algorithms to find spatial patterns with temporal features and temporal patterns with spatial features. We applied these algorithms to a shared fMRI repository of face recognition tasks. We matched the classes found and our results of functional connectivity analysis with partitioning of BOLD signal signatures. We compared the outcomes using the just acquired model-based knowledge as likely ground truth, confirming the role of Fusiform Brain Regions. In general, partitioning results show a better spatial clustering than temporal clustering for both algorithms. In the case of temporal clustering, FCM outperforms Neural Gas. The relevance of brain subregions related to face recognition were correctly distinguished by the algorithms and the results are in agreement with the current neuroscientific literature.
Classification of human emotions remains an important and challenging task for many computer vision algorithms, especially in the era of humanoid robots which coexist with humans in their everyday life. Currently proposed methods for emotion recognition solve this task using multi-layered convolutional networks that do not explicitly infer any facial features in the classification phase. In this work, we postulate a fundamentally different approach to solve emotion recognition task that relies on incorporating facial landmarks as a part of the classification loss function. To that end, we extend a recently proposed Deep Alignment Network (DAN), that achieves state-of-the-art results in the recent facial landmark recognition challenge, with a term related to facial features. Thanks to this simple modification, our model called EmotionalDAN is able to outperform state-of-the-art emotion classification methods on two challenging benchmark dataset by up to 5%.
Bertrand proposed the notion of a P-simple set for constructing topology-preserving reductions. In this paper, we define the maximalness of a P-simple set, give a new sufficient condition for topology-preserving reductions acting on (8,4) pictures on the square grid, and it is proved that this condition designates a maximal P-simple set.
Thinning is an iterative object reduction: border points that satisfy some topological and geometric constraints are deleted until stability is reached. If a border point is not deleted in an iteration, conventional implementations take it into consideration again in the next step. With the help of the concepts of a 2D-simplifier point and a weak-3D-simplifier point, rechecking of some 'survival' points is not needed. In this work an implementation scheme is reported for sequential thinning algorithms, and it is shown that the proposed method can be twice as fast as the conventional approach in the 2D case.
In robotic applications, highly specific objects such as industrial parts, for example, often need to be recognized. In these cases methods can’t rely on the online availability of large labeled training data sets or pre-trained models. This is especially true for depth data, thus making it challenging for deep learning (DL) approaches. Therefore, this work analyzes the performance of various traditional (global or part-based) and DL features on a restricted depth data set, depending on the tasks complexity. While the sample size is small, we can conclude that pre-trained DL descriptors are the most descriptive, but not by a statistically significant margin and therefore part-based descriptors are still a viable option for small, but difficult 3D data sets.
Imaging biomarkers are becoming important in both research and clinical studies. This study is focused on developing measures of tumour mean, fractal dimension, homogeneity, energy, skewness and kurtosis that reflect the values of the pharmacokinetic (PK) parameters within the breast tumours, evaluate those using clinical data, and investigate their feasibility as a biomarker to discriminate malign from benign breast lesions. In total, 75 patients with breast cancer underwent Dynamic Contrast Enhanced-Magnetic Resonance Imaging (DCE-MRI). Axial bilateral images with fat-saturation and full breast coverage were performed at 3T Siemens with a 3D gradient echo-based TWIST sequence. The whole tumour mean, fractal dimension, homogeneity, energy, skewness and kurtosis of $$K^{trans}$$ and $$V_e$$ values were calculated. Median of both the mean and the fractal dimension of $$K^{trans}$$ and $$V_e$$ for benign and malignant tumour show significant discrimination. Further, the median of skewness and kurtosis of $$V_e$$ significantly vary between benign and malignant cases. In conclusion, the mean and the fractal dimension of both $$K^{trans}$$ and $$V_e$$ and skewness and kurtosis of $$V_e$$ for typical breast cancer, computed from PK parametric maps, show potential as a biomarker for breast tumour diagnosis either as a benign or malignant.
A new syntactic model, called pure hexagonal context free grammar is introduced based on the notion of pure two-dimensional context-free grammar. These grammars generate hexagonal picture arrays on triangular grids. We also examine certain closure properties of pure hexagonal context free languages.
Mammogram images are broadly categorized into two types: carniocaudal (CC) view and mediolateral oblique (MLO) view. In this paper, we study the effect of different image views for mammogram mass classification. For the experiments, we consider a dataset of 328 CC view images and 334 MLO view images (almost equal ratio) from a publicly available film mammogram image dataset [3]. First, features are extracted using a novel radon-wavelet based image descriptor. Then an extreme learning machine (ELM) based classification technique is applied and the performance of five different ELM kernels are compared: sigmoidal, sine, triangular basis, hard limiter and radial basis function. Performances are reported in terms of three important statistical measures namely, sensitivity or true positive rate (TPR), specificity or false negative rate (SPC) and recognition accuracy (ACC). Our experimental outcome for the present setup is two-fold: (i) CC view performs better then MLO for mammogram mass classification, (ii) hard limiter is the best ELM kernel for this problem.
We propose a reconstruction method for a theoretic projection acquisition technique, where we assume that the object of study consists of a finite number of materials, and we can separately measure the amount of materials along the paths of projection beams. The measurement decomposes the projections for separating materials, i.e., we get a separate projection set for each material (called decomposed projections), and each projection set holds information on one material only. We describe a mathematical formulation where the newly proposed reconstruction problem is formalised by an equation system and show that the model can be solved by equation system-based reconstruction techniques like the SIRT method while maintaining convergence. We test the theoretic setup on simulated data by reconstructing phantom images from simulated projections and compare the results to reconstructions from classical X-ray projections. We show that using decomposed projections can lead to better results from 20 times less number of projections than the classical X-Ray tomography.
Binary tomography reconstructs binary images from a low number of their projections. Often, there is a freedom how these projections can be chosen which can significantly affect the quality of reconstructions. We apply sequential feature selection methods to find the ‘most informative’ projection set based on a blueprint image. Using various software phantom images, we show that these methods outperform the previously published projection selection algorithms.
We propose a fast object detector, based on Convolutional Neural Network (CNN). The object detector, which operates on RGB images, is designed for a mobile robot equipped with a robotic manipulator. The proposed detector is designed to quickly and accurately detect objects which are common in small manufactories and workshops. We propose a fully convolutional architecture of neural network which allows the full GPU implementation. We provide results obtained on our custom dataset based on ImageNet and other common datasets, like COCO or PascalVOC. We also compare the proposed method with other state of the art object detectors.
Radiotherapy is one of the most common methods to treat different cancer cells in clinical application despite having harmful effects on healthy tissues. Radiobiological experiments are very important to determine the irradiation-caused acute and chronic effects to define the exact consequences of different irradiation sources. Photon irradiation has been used on zebrafish embryos, a very new in vivo and appropriate model system in radiobiology. After irradiation, dose-dependent morphological changes were observable in the embryos. These morphological deteriorations were measured manually by biologist researchers during three weeks, which was an extremely time demanding process (15 min per image). The aim of this project was to automate this evaluating process, to save time for researchers and to keep the consistence and accuracy of the evaluation. Hence, an algorithm was developed and used to detect the abnormal development of zebrafish embryos.
This contribution presents a method for numerical analysis of solids whose boundaries are represented by oriented point clouds. In contrast to standard finite elements that require a boundary-conforming discretization of the domain of interest, our approach works directly on the point cloud representation of the geometry. This is achieved by combining the inside-outside information that is inferred from the members of the point cloud with a high order immersed boundary technique. This allows for avoiding the challenging task of surface fitting and mesh generation, simplifying the image-based analysis pipeline drastically. We demonstrate by a numerical example how the proposed method can be applied in the context of linear elastostatic analysis of solids.
In image-based coin detection, making the image readable is an indispensable part of the feature extraction. However using a 2-D image processing approach for detecting a counterfeit coin is nearly impossible in case of destroyed coins whose textures are severely burnt, sulfated, rusted, or colored. In this research, we used a 3-D scanner to scan and model an acceptable number of coins capturing height and depth instead of levels of color. The most important advantage of 3-D scanning is to compensate for the above-mentioned destructions of the coin surface. Despite this advantage, we had several unexpected degradations due to shiny coin images. To solve this problem, the 3-D image was decomposed column-wise to a number of separate 1-D signals, which were analyzed separately and restored by the proposed method. This approach gave remarkable results when used to extract valuable features.
Shape classification is a required task in many systems for image and video understanding. Implicit shape representations, such as the solutions to the Eikonal or Poisson equations defined on the shape, have been shown to be particularly effective for generating features that are useful for classification. The Poisson-based shape representation can be derived at each point inside the shape as the expected time for a particle undergoing Brownian motion to hit the shape boundary. This representation has no natural generalization when considering points outside of a shape, however, because the corresponding Brownian motion would have infinite expected hitting time. In this article, we modify the Brownian motion model by introducing an exponential lifetime for the particle, yielding a random variable whose expected value satisfies a screened Poisson equation that can be solved at points both interior and exterior to the shape. We then show how moments of this new random variable can be used to improve classification results on experiments with natural silhouettes and handwritten numerals.
Given an input 3D image, in this paper we first segment it into several clusters by extending the 2D harmonic edge-weighted centroidal Voronoi tessellation (HEWCVT) method to the 3D image domain. The Dual Contouring method is then applied to construct tetrahedral meshes by analyzing both material change edges and interior edges. An anisotropic Giaquinta-Hildebrandt operator (GHO) based geometric flow method is developed to smooth the surface with both volume and surface features preserved. Optimization based smoothing and topological optimizations are also applied to improve the quality of tetrahedral meshes. We have verified our algorithms by applying them to several datasets.
