9th International Conference on Pattern Recognition Systems (ICPRS 2018)
Buy conference proceeding
- Location: Valparaíso, Chile
- Conference date: 22-24 May 2018
- ISBN: 978-1-78561-887-1
- Conference number: CP745
- The following topics are dealt with: medical image processing; image classification; object detection; image segmentation; computer vision; text analysis; feature extraction; neural nets; positron emission tomography; image denoising.
17 items found
-
3D Multi-Scale Convolution Nets for Pulmonary Nodule Detection
- Author(s): Rundong Wang ; Yuancheng Wang ; Yuhao Zhang ; Qiao Wang
- + Show details - Hide details
-
p.
(6 pp.)
Pulmonary nodule detection is a challenging task, particularly in early diagnosis with tiny nodules. The main concerns to detector design are involved with the resolution of CT images, the high variance of nodule sizes and contextual reasoning. In this paper, we present a novel multi-scale improving strategy which can well decrease the number of undetected tiny pulmonary nodules. More specifically, we first implement and train the top-down module-based 3D Regional Proposal Network(RPN) as the a “one-size-fits-all” baseline approach. Then, we propose a simple and efficient multi-scale approach in training and testing within the same network architecture. Finally, we evaluate the performance of these two systems and find that the multi-scale method exceeds the baseline one, specially the task of tiny pulmonary nodule detection.
-
Convolutional Neural Networks Applied to Multiple Sclerosis Lesion Segmentation on 3D Brain Magnetic Resonance Images
- Author(s): R. Naranjo ; G. Ulloa ; H. Allende-Cid ; H. Allende
- + Show details - Hide details
-
p.
(5 pp.)
Multiple Sclerosis (MS) is a disabling disease which affects the central nervous system. The segmentation of the multiple sclerosis lesions in 3D brain Magnetic Resonance (MR) images is a fundamental task in diagnosis and tracking of this disease. The process of segmentation of the lesions is usually performed manually by experts, however, there exists interest in the automation of this task in order to speed up and standardize this process. To this end, multiple automated segmentation techniques have been proposed to effectively detect MS lesions. In this work, the performance of Convolutional Neural Networks (CNN) applied to the problem of MS lesion detection in 3D brain MR images will be compared to other state of art proposals.
-
Dynamic PET image denoising
- Author(s): P. Gonzalez ; B. Alcaino ; R. Barrientos ; M. Mora ; F. Tirado ; C. Tauber
- + Show details - Hide details
-
p.
(6 pp.)
Dynamic Positron Emission Tomography (dPET) images are inherently affected by noise and low spatial resolution. The problems aforementioned may lead to incorrect estimation of the uptake of the tracer in tissues. In this work, we present a novel method for enhancing the signal-to-noise ratio of dPET images. The method consist in a edge preserving filter based upon an indirect image. The indirect image give orientation to the treatment so as to process all frames at the same fashion. We exploit the spatial and temporal information along the entire sequence in order to adapt the filtering process to preserve edges between functional regions. Comparative experimentations on realistic simulations validate the potential of the proposed method.
-
Dynamic PET reconstruction using KIBF 4D filter within reconstruction algorithm
- Author(s): R. Delaplace ; S. Stute ; C. Tauber
- + Show details - Hide details
-
p.
(4 pp.)
Positron emission tomography (PET) are noisy, especially when working on dynamic acquisitions. To contest that noise, we developed a method that combine the EM algorithm, the most usual reconstruction algorithm, and dynamic filters. We compared our method to OSEM reconstruction and post filtering alone using three different quantitative criteria. Those criteria are Signal-to-Noise Ratio (SNR), bias and Pratt's Figure of Merit (PFOM). We demonstrate that our method provides better performances according to this criteria, soit is usable for both research and clinic.
-
Detection of People Boarding/Alighting a Metropolitan Train using Computer Vision
- Author(s): M. Belloc ; S.A. Velastin ; R. Fernandez ; M. Jara
- + Show details - Hide details
-
p.
(6 pp.)
Pedestrian detection and tracking have seen a major progress in the last two decades. Nevertheless there are always application areas which either require further improvement or that have not been sufficiently explored or where production level performance (accuracy and computing efficiency) has not been demonstrated. One such area is that of pedestrian monitoring and counting in metropolitan railways platforms. In this paper we first present a new partly annotated dataset of a full-size laboratory observation of people boarding and alighting from a public transport vehicle. We then present baseline results for automatic detection of such passengers, based on computer vision, that could open the way to compute variables of interest to traffic engineers and vehicle designers such as counts and flows and how they are related to vehicle and platform layout.
-
Unsupervised Band Selection in Hyperspectral Images using Autoencoder
- Author(s): M. Habermann ; V. Fremont ; E.H. Shiguemori
- + Show details - Hide details
-
p.
(6 pp.)
Hyperspectral images provide fine details of the observed scene from the exploitation of contiguous spectral bands. However, the high dimensionality of hyperspectral images causes a heavy burden on processing. Therefore, a common practice that has been largely adopted is the selection of bands before processing. Thus, in this work, a new unsupervised approach for band selection based on autoencoders is proposed. During the training phase of the autoencoder, the input data samples have some of their features turned to zero, through a masking noise transform. The subsequent reconstruction error is assigned to the indices with masking noise. The bigger the error, the greater the importance of the masked features. The errors are then summed up during the whole training phase. At the end, the bands corresponding to the biggest indices are selected. A comparison with four other band selection approaches reveals that the proposed method yields better results in some specific cases and similar results in other situations.
-
Bees Algorithm for Efficiently Separating n-Dimensional Polynomials
- Author(s): M. Alrajab ; H. Aljabbouli ; J.-C. Ngatchou
- + Show details - Hide details
-
p.
(5 pp.)
Separating n-dimensional all-zero polynomials has many important applications in many disciplines. This paper proposes a new method to approximate the one-dimensional coefficients of the n-dimensional all-zero. The proposed solution is based on the Bees algorithm which is a powerful technique for multi-dimensional optimization problems with local and global search. The results presented in this paper show significant improvements when compared to previous methods including: Neural Networks and Genetics Algorithm. Testing was performed on three different polynomials to demonstrate the efficiency of the proposed solution.
-
A multi-directional Gradient with Bi-Geometric Calculus to Detect Contours in Images with Multiplicative Noise
- Author(s): M. Acevedo-Letelier ; K. Vilches ; M. Mora
- + Show details - Hide details
-
p.
(7 pp.)
In this paper a new operator is presented for the detection of contours in images with multiplicative noise, by using the operations introduced in the bi-geometric calculus, since recent results in the literature show that multiplicative operators tend to make more accurate approximations of the reality in images with multiplicative noise. The operator introduced corresponds to a multiplicative multi-directional gradient. The Global Efficiency was used as performance function to make a comparison about the effectiveness in the detection of contours, between the multi-gradient and its multiplicative version. These operators are applied on some images (one synthetic and another real), under a threshold for each noise level, then the function of optimal performance is obtained over a continuous range of noise, and thus an objective comparison between both operators is presented. According to the results obtained from the objective comparison, the multiplicative multi-directional gradient operator presents improved efficiency in obtaining contours versus its classical version.
-
Boosting Text Clustering using Topic Selection
- Author(s): M. Mendoza ; P. Ormeno ; C. Valle
- + Show details - Hide details
-
p.
(6 pp.)
Latent Dirichlet Allocation (LDA) is a key topic modeling algorithm in the text mining field. Despite the great success of LDA, the state of the art reports that LDA is sensitive to the choice of hyper parameters and accordingly, the quality of the topics found depends on tuning. Instead of looking for the optimal hyper parameters of LDA for a given corpus, we propose a strategy for topic selection and aggregation that exploits hyper parameter variability, as the number of topics inferred, to boost the quality of the topics found. We show that our approach is simple and very effective to boost topic models. Experimental results show that our proposal improves the quality of the topics found, favoring document and term clustering tasks.
-
A dataset of Warped Historical Arabic Documents
- Author(s): A. Dulla
- + Show details - Hide details
-
p.
(6 pp.)
Historical documents are considered one of the most important human wealth and a source of intellectual production. Unfortunately, due to aging effects, multiple noises and arbitrary geometric distortions are found in the document image. This paper presents a novel dataset (and the methodology used to create it) based on an extensive variety of historical Arabic documents containing clean information basic and homogeneous-page layouts. The tests are executed on printed and handwritten documents acquired respectively from some imperative libraries, for example, Qatar Digital Library, the British Library and the Library of Congress. We have collected and commented on 200 archival document images from various sites and time periods. It is based on different documents from the 16th19th century. The dataset involves varying page layouts and degradations that challenge text line segmentation techniques. Ground truth is created using the Aletheia tool by PRImA and stored in an XML representation, in the PAGE format. The dataset offered will be effectively accessible to specialists worldwide for inquire about into the impediments confronting different historical Arabic documents, for example, geometric correction of historical Arabic documents.
-
Automatic Opinion Classification Using Conformal Predictors
- Author(s): G. Farias ; J. Leon ; E. Fabregas ; S. Dormido-Canto
- + Show details - Hide details
-
p.
(7 pp.)
This paper describes the use of conformal predictors to collect reliable positive or negative opinions of a topic from twitter users. Conformal predictors provides not only the label of sample, but also the reliability of this prediction. This feature of conformal algorithms can be naturally used to consider only opinions with high credibility of a specific topic in order to follow a recommendation. This approach can be easily extended to almost any kind of topic, the only one condition is to have previously labelled opinions of the topic that can be used for training or building the conformal predictor. The approach is tested with opinions of Spanish twitter users that recommend (or not) to watch a movie.
-
Semi-supervised regression based on tree SOMs for predicting students performance
- Author(s): H. Nuñez ; G. Maldonado ; C.A. Astudillo
- + Show details - Hide details
-
p.
(7 pp.)
The Sistema de Medición de la Calidad de la Educación (Education Quality Measurement System, SIMCE) is an annual survey designed to evaluate the Chilean educational system through standard tests that measure the abilities and knowledge of the students. The results of the SIMCE exams provide important information for analyzing the learning processes of the students. Additionally, these results allow the identification of strengths and weaknesses for the elaboration of public policies. In this paper, we design a semi-supervised regressor called TTOSOM as Regression Model (TTOSOM-RM), that inherits the properties of the neural network called Tree-based Topology Oriented SOM (TTOSOM) [1]. The goal is to predict the performance of fourth grade students of the Chilean educational system in the SIMCE test. The proposed model successfully predicts the SIMCE scores, producing a lower absolute mean error when compared with other state-of-the-art methods.
-
An Enhanced Algorithmic Approach for Automatic Defects Detection in Green Coffee Beans
- Author(s): C.E. Portugal Zambrano ; J.C. Gutierrez Caceres ; J. Ramirez Ticona ; N.J. Beltran-Castanon ; J.M. Ramos Cutipa ; C.A. Beltran-Castañon
- + Show details - Hide details
-
p.
(8 pp.)
Classification green coffee beans is one of the main tasks during the quality grading process. This evaluation is normally carried out by specialist doing a visual inspection or using traditional instruments which have some limitations. This work is focused on the implementation of a computer vision system combining a hardware prototype and a software module. The hardware was made to guarantee the controlled conditions to capture the images of green coffee beans, the software is based on computer vision algorithms in order to detect defects of the coffee beans. The novelty of our proposal is the combination of algorithms to enhance the accuracy and the high number of defects detected. We applied a White Patch algorithm as an image enhancement procedure, color histograms as feature extractor and Support Vector Machine (SVM) for the classification task. It was constituted an image beans database of 1930 instances, and it was extracted 768 features, finally, the model was applied over 13 categories of defects described by the Specialty Coffee Association of America (SCAA). Results of classification achieved a 98.8% of overall accuracy detection, therefore the proposed system proved to be effective in classifying physical defects of green coffee beans. With this work we showed that the grading green coffee process can be automatized, adding a new paradigm in quality evaluation task to enhance the coffee industry.
-
Silhouette-based human action recognition with a multi-class support vector machine
- Author(s): L. Gonzalez ; S.A. Velastin ; G. Acuna
- + Show details - Hide details
-
p.
(5 pp.)
Computer vision systems have become increasingly popular, being used to solve a wide range of problems. In this paper, a computer vision algorithm with a support vector machine (SVM) classifier is presented. The work focuses on the recognition of human actions through computer vision, using a multi-camera dataset of human actions called MuHAVi. The algorithm uses a method to extract features, based on silhouettes. The challenge is that in MuHAVi these silhouettes are noisy and in many cases include shadows. As there are many actions that need to be recognised, we take a multiclass classification approach that combines binary SVM classifiers. The results are compared with previous results on the same dataset and show a significant improvement, especially for recognising actions on a different view, obtaining overall accuracy of 85.5% and of 93.5% for leave-one-camera-out and leave-one-actor-out tests respectively.
-
Towards landmarks prediction with Deep Network
- Author(s): Van Linh Le ; M. Beurton-Aimar ; A. Zemmari ; N. Parisey
- + Show details - Hide details
-
p.
(6 pp.)
Morphometry landmarks are used in many biological applications. Mostly, the landmarks are defined manually or semiautomatically by applying image processing techniques. In recent years, Deep Learning is known as a solution to achieve image analysis tasks such as classification, recognition, or face detection. In this context, we present a Convolutional Neural Network (CNN) model to predict landmarks on 2D anatomical images, specifically beetle's images. The dataset includes the images of collecting from 293 beetles. For each beetle, 5 images are available corresponding to head, pronotum and body parts. For each part, a set of manual landmarks has been positioned by an entomologist. In this work, we have focused on prediction of pronotum landmarks. The proposed CNN model is designed from an elementary block of three layers: convolution, pooling, and dropout. The network is trained in two different ways: from scratch or after a step of fine-tuning. The fine-tuning parameters are obtained by training on all parts of beetles before to be applied to the pronotum. The quality of predicted landmarks is evaluated by calculating the distance in pixels between the coordinates of the predicted and manual landmarks which are considered as the ground truth. The obtained results by applying fine-tuning steps are considered to be statistically good enough to replace the manual ones for the different morphometry analysis.
-
Motorcycle detection and classification in urban Scenarios using a model based on Faster R-CNN
- Author(s): J.E. Espinosa ; S.A. Velastin ; J.W. Branch
- + Show details - Hide details
-
p.
(6 pp.)
This paper introduces a Deep Learning Convolutional Neutral Network model based on Faster-RCNN for motorcycle detection and classification on urban environments. The model is evaluated in occluded scenarios where more than 60% of the vehicles present a degree of occlusion. For training and evaluation, we introduce a new dataset of 7500 annotated images, captured under real traffic scenes, using a drone mounted camera. Several tests were carried out to design the network, achieving promising results of 75% in average precision (AP), even with the high number of occluded motorbikes, the low angle of capture and the moving camera. The model is also evaluated on low occlusions datasets, reaching results of up to 92% in AP.
-
Considerations on anisotropic models of image processing based on partial differential equations
- Author(s): D.A. Pulido and J.J. Rodriguez
- + Show details - Hide details
-
p.
(6 pp.)
Multiscale analysis methods suggest the use of partial differential equations that involve curvature to decrease the alteration of edges and achieve a better smoothing in internal regions of an image, for which the experimental results are presented using these new equations and the combination of these with the Perona-Malik model, showing and comparing their experimental results.
