10th International Conference on Pattern Recognition Systems
Buy conference proceeding
- Location: Tours, France
- Conference date: 8-10 July 2019
- ISBN: 978-1-83953-108-8
- Conference number: CP761
- The following topics are dealt with: image classification; learning (artificial intelligence); convolutional neural nets; object detection; medical image processing; feature extraction; computer vision; brain; biomedical MRI; image segmentation.
19 items found
-
Defect Detection in Tunnel Images using Random Forests and Deep Learning
- Author(s): G. Decor ; M.D. Bah ; P. Foucher ; P. Charbonnier ; F. Heitz
- + Show details - Hide details
-
p.
(NaN)
Tunnel maintenance requires a complicated and constraining visual inspection. In order to automate this task, we propose to evaluate and compare three statistical learning algorithms, a random forest and two convolutional networks, dedicated to the detection of defects (e.g. cracks) on tunnel linings. Each model is trained on datasets of our own, consisting of images of concrete walls and masonry walls. We show that these learning-based approaches are competitive with the state of the art on this application domain.
-
Deep Convolutional Neural Network to predict 1p19q co-deletion and IDH1 mutation status from MRI in Low Grade Gliomas
- Author(s): S.R. González ; I. Zemmoura ; C. Tauber
- + Show details - Hide details
-
p.
(NaN)
Predicting arm-chromosomes 1p19q co-deletion and IDH1 mutation in Low Grade Gliomas is determinant in the treatment planning and follow up of the patients. This study aims at proposing a non-invasive method, based on multimodal MR images using convolutional neural networks. The proposed approach consists in several preprocessing steps and an Inception architecture. We present comparative results on a publicly available dataset. The proposed Inception v3 architecture obtain a F1-score of 91.38 ± 5.7% in the test set, when classifying images between 1p19q preserved and codeleted and 82.07% ± 12% when classifying images between with and without IDH1 mutation.
-
Unsupervised classification of street architectures based on InfoGAN
- Author(s): Ning Wang ; Xianhan Zeng ; Renjie Xie ; Zefei Gao ; Yi Zheng ; Ziran Liao ; Junyan Yang ; Qiao Wang
- + Show details - Hide details
-
p.
(NaN)
Street architectures play an essential role in city image and streetscape analysing. However, existing approaches are all supervised which require costly labeled data. To solve this, we propose a street architectural unsupervised classification framework based on Information maximizing Generative Adversarial Nets (InfoGAN), in which we utilize the auxiliary distribution Q of InfoGAN as an unsupervised classifier. Experiments on database of true street view images in Nanjing, China validate the practicality and accuracy of our framework. Furthermore, we draw a series of heuristic conclusions from the intrinsic information hidden in true images. These conclusions will assist planners to know the architectural categories better.
-
3D-SiameseNet to Analyze Brain MRI
- Author(s): C. Ostertag ; M. Beurton-Aimar ; T. Urruty
- + Show details - Hide details
-
p.
(NaN)
Data used in preparation of this article were obtained from the Alzheimers Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.usc.edu/wpcontent/uploads/how to apply/ADNI Acknowledgement List.pdf.
-
Predicting air quality using deep learning in Talca City, Chile
- Author(s): C.A. Astudillo ; L. González-Martínez ; E. Zapata-González
- + Show details - Hide details
-
p.
(NaN)
Nowadays, during the colder season, a number of cities in central and southern Chile are affected by a problem of air pollution. This problem is associated to the topography of the central valley, high emissions of domestic wood-burning stoves, and other meteorological characteristics such as the lack of rain or when the prevailing winds are calmed. Although, numerous studies have been undertaken for predicting the air quality in Santiago city, other medium-sized cities have seldom been studied. The present work focuses on the problem of air quality in Talca city (35°26'S; 71°44'W), a medium sized city located in Central Chile. The objective of the study is to predict, with a day in advance, the particulate matter with a diameter less than or equal to 2.5 micrometers (PM2.5). For this purpose we have used a deep learning neural network. The algorithm learns from historical records of air quality pollution as well as meteorological information at three monitoring stations. Unlike state-of-the-art methods that require intensive computational power to simulate the weather conditions, our proposed solution uses only pollutants measures of the stations in the city. We use exactly 24 records for a particular day, one for each hour. Our study focuses on the autumn-winter season for 3 years, including data for 612 days, i.e., 151 days per year (from April 1st, until August 31st). Our results prove the high capacity of RNN as a predictor algorithm for environmental emergency episodes in the three monitoring stations of the city. As a result, the model is an alternative for local authorities because it would improve the current forecast system of the city.
-
QLTL: a Simple yet Efficient Algorithm for Semi-Supervised Transfer Learning
- Author(s): B. Muller and R. Lengelle
- + Show details - Hide details
-
p.
(NaN)
Most machine learning techniques rely on the assumption that training and target data share a similar underlying distribution. When this assumption is violated, they usually fail to generalise; this is one of the situations tackled by transfer learning: achieving good classification performances on different-butrelated datasets. In this paper, we consider the specific case where the task is unique, and where the training set(s) and the target set share a similar-but-different underlying distribution. Our method, QLTL: Quadratic Loss Transfer Learning, constitutes semi-supervised learning: we train a set of classifiers on the available training data in order to input knowledge, and we use a centred kernel polarisation criterion as a way to correct the density probability function shift between training and target data. Our method results in a convex problem, leading to an analytic solution. We show encouraging results on a toy example with covariate shift, and good performances on a text-document classification task, relatively to recent algorithms.
-
Minimum volume Constrained non-negative matrix factorization applied to the monitoring of active cosmetic ingredient into the skin in Raman imaging
- Author(s): A. Stella ; F. Bonnier ; L. Miloudi ; A. Tfayli ; F. Yvergnaux ; E. Munnier ; C. Tauber
- + Show details - Hide details
-
p.
(NaN)
The hyperspectral imaging is commonly used in the cosmetic area. Indeed, spectroscopic imaging techniques are usually employed to study the molecular composition of a cosmetic product, notably the Raman imaging. For this matter, Nonnegative Constrained Least Square (NCLS), has been studied previously and has provided accurate distribution maps of Active Cosmetic Ingredient (ACI) with its associated penetration profile. However, it remains a supervised method since it requires an a priori knowledge of the Raman fingerprint of the ACI to track it and the availability of a large number of spectra from control data affects its performance. This work presents the comparison of a Minimum Volume Constrained Non-negative Matrix Factorization (MVC-NMF) with the NCLS and a popular method in the chemometry community, Multivariate Curve Resolution Alternating Least Square (MCR-ALS) for hyperspectral image analysis. MVC-NMF proposes an unsupervised geometric approach to better fit a linear model to the data that provides lower modelling residuals. We also evaluate the parameter selection of the right number of constituent of Raman imaging from skin samples. It is shown that the MVC-NMF was able to accurately estimate the Raman spectrum of the ACI without supervision.
-
A New Normalized Method of Object Shape-based Recognition and Localization
- Author(s): B. Moradi ; H. Abdulrahman ; B. Magnier
- + Show details - Hide details
-
p.
(NaN)
This paper introduces a new normalized measure for the assessment of a contour-based object pose. This algorithm allows a supervised assessment of recognition and localization of known objects, the differences between a reference edge map and a candidate image are quantified by computing a performance measure. This measure is normalized and is able determine the degree to which an object shape differs from a desired position. Compared to 6 other approaches, experiments on real images at different sizes/scales exhibit the suitability of the new method for an object pose or shape matching estimation.
-
Face segmentation based object localisation with deep learning from unconstrained images
- Author(s): T. Alzahrani and W. Al-Nuaimy
- + Show details - Hide details
-
p.
(NaN)
Localisation and segmentation of the face region is a key step and prerequisite in automated face recognition systems. Over the past few decades, there has been particular interest and efforts in finding automated, robust, accurate and reliable face detection methods. In this paper, we propose a deep learning-based method to segment face boundaries from unconstrained images. The proposed deep learning model defines the detection of face boundary formula as a combined object localisation and segmentation task. Incorporating an object localisation in a framework developed for segmentation purpose aims to improve the segmentation accuracy achieved by a fully convolutional network. The proposed model is trained on 2800 in-the-wild face images and tested on 700 images achieving dice coefficient of 95.67%, Jaccard score of 91.89%, accuracy of 98.18%, sensitivity of 96.14% and specificity of 98.76%. The experimental results demonstrate that the proposed deep learning method outperforms most of state of the art methods in automatic face detection and segmentation on a challenging dataset.
-
Classification of Mouth Gestures in German Sign Language using 3D Convolutional Neural Networks
- Author(s): N. Wilson ; M. Brumm ; R.-R. Grigat
- + Show details - Hide details
-
p.
(NaN)
Automatic recognition of sign language gestures is becoming necessary with an increased interest into human-computer interaction in sign language as well as automatic translation from sign language. Most of the research on sign language recognition focuses on hand gesture recognition. However, there are also non-manual signals in sign language. Mouth gestures represent mouth shapes that add information to the hand gestures not related to spoken language visemes. For German Sign Language, mouth gesture recognition would be an important addition to manual gesture recognition. This research work evaluates the method 3D convolutional neural networks for recognising mouth gestures in German Sign Language. For the recognition of certain mouth gestures, temporal information is mandatory and the extraction of both spatial and temporal features by 3D convolutional networks makes the classification of all gestures easier. Our research work compares how different initialisations affect learning and classification by the network. We achieve an accuracy of around 68% on testing 10 classes of mouth gestures in German Sign Language.
-
Simultaneous filtering and sharpening of vector-valued images with numerical schemes
- Author(s): P. Gonzalez ; T. Garcia ; P. Spiterif ; C. Tauber
- + Show details - Hide details
-
p.
(NaN)
The present work deals with the development of a novel approach to increase the quality of 3D vector-valued (4D) images. We model the problem as a coupled anisotropic convection diffusion filtering scheme based upon local structure analysis. The orientation of the denoising and shock filtering are based upon an analysis of the local structure of the vector-valued image through a tensor that accounts for the distribution of the noise between the channels. With this approach, the noise is reduced in homogeneous regions while the vector edges are sharpened orthogonally to the flow of diffusion. We present the continuous problem and a corresponding discretization in space and time by a numerical way. Results on realistic dynamic PET image simulations illustrate the potential of the proposed method, which led to distinct improvements of figures of merit over several other approaches from the literature.
-
Multiplicative Local Binary Patterns (MuLBP)
- Author(s): M. Mora ; M. Silva-Ibarra ; M. Acevedo-Letelier
- + Show details - Hide details
-
p.
(NaN)
Speckle is a multiplicative noise that greatly deteriorates images. In this paper a model of Local Binary Patterns (LBP) adapted to images with speckle (MuLBP) is proposed. The multiplicative model is constructed by substituting the additive comparisons of the traditional LBP for multiplicative comparisons from the Bigeometric Calculus. The experiments were carried out considering the 10.824 images of the KTH-TIPS2, FMD, CASIA and UFI databases. To compare the additive and multiplicative models, the Euclidean distance between the LBP histograms of the image with noise and without noise is adopted. The results indicate that, the distance between the histograms, the image with noise with respect to the image without noise, is smaller for the multiplicative models than for the traditional additive models. The above means that, the multiplicative LBP represent in a better way the textures in images contaminated with speckle.
-
A Novel Baseline Estimation Technique for Geometric Correction of Historical Arabic Documents Based on Voronoi Diagrams
- Author(s): A. Dulla and A. Antonacopoulos
- + Show details - Hide details
-
p.
(NaN)
Since Arabic writing has a robust baseline, several state-of-the-art recognition systems for handwritten Arabic produce use of baseline-dependent characteristics. For modern Arabic documents, the baseline can be detected reliably by obtaining the maximum in the horizontal projection profile or the Hough transformed image. However, the performance of these techniques leaks significantly on Historical Arabic Documents. In this paper, we introduce an effective novel approach to baseline detection in Historical Arabic Documents which is based on Based on Voronoi Diagrams. The proposed technique is carried out, verified and validated on a dataset of Warped Historical Arabic Documents based on affecting by warping percentage.
-
Structural information and (hyper)graph matching for MRI piglet brain extraction
- Author(s): A. Durandeau ; J.-B. Fasquel ; I. Bloch ; E. Mazerand ; P. Menei ; C. Montero-Menei ; M. Dinomais
- + Show details - Hide details
-
p.
(NaN)
In the context of the study of the maturation process of the infant brain, this paper focuses on postnatal piglet brain, whose structure is similar to the one of an infant. Due to the small size of the piglet brain and the abundance of surrounding fat and muscles, the automatic brain extraction using correctely initialized deformable models is tedious, and the standard approach used for human brain does not apply. The paper proposes an original brain extraction method based on a deformable model, whose initialization is guided by a priori known relationships between some anatomical structures of the head. This concerns a structural model related to a priori known inclusion and photometric relationships between eyes, nose and other internal head entities (fat and muscles). This a priori structural information also involves the relative position of both eyes and nose, assumed to be an anatomical invariant similar to a triangle. Using this structural model, our proposal detects both eyes and nose, from which one deduces the brain center, for finally initializing deformable models. Anatomical structures are retrieved by matching observed relationships with those embedded in the a priori structural model. This involves graph and hypergraph matching, where hypergraph matching concerns relative position of eyes and nose (ternary constraint related to these 3 entities). The method has been implemented and preliminary experiments have been performed on a set of 6 piglets, to evaluate the accuracy of the brain center localization, the one of the final brain extraction using deformable models. The brain center is correctly localized with a mean error of 1.7 mm, underlying the relevance of the approach. The mean similarity index has been measured to be of 0.85 (with a standard deviation of 0.04). More generally, this work illustrates the potential of considering high level a priori known relationships, related to anatomical invariants, managed using graph and hypergraph matching.
-
ParKerC: Toolbox for Parallel Kernel Clustering Methods
- Author(s): S. Mouysset and R. Guivarch
- + Show details - Hide details
-
p.
(NaN)
A large variety of fields such as biology, information retrieval, image segmentation needs unsupervised methods able to gather data without a priori information on shapes or locality. By investigating a parallel strategy based on overlapping domain decomposition, we present a toolbox which is a parallel implementation of two fully unsupervised kernel methods respectively based on density-based properties and spectral properties in order to treat large data sets in fields of pattern recognition.
-
Classification of Eimeria species from digital micrographies using CNNs
- Author(s): D.F. Monge and C.A. Beltran
- + Show details - Hide details
-
p.
(NaN)
This paper presents a model for the classification of the seven species of avian Eimeria, the protozoan parasite that causes avian coccidiosis. Digital micrographs dataset consists of 4485 isolated samples of the various species of oocytes (status of the Eimeria protozoon in which the internal structure is visually different in each species). The proposed solution applied a convolutional neural network architecture for the classification of the oocytes. Different experiments were developed to enhance the previous results of the literature, and with our proposal, we obtained a better average of correct classification for the seven species, reaching 90.42% of precision. Finally, with our strategy we used for the first time a CNN model over the Eimeria dataset, demonstrating that CNN is a robust technique for artificial vision problems.
-
A Benchmark Dataset for RGB-D Sphere Based Calibration
- Author(s): D.J.T. Boas ; S. Poltaretskyi ; J.-Y. Ramel ; J. Chaoui ; J. Berhouet ; M. Slimane
- + Show details - Hide details
-
p.
(NaN)
Accurate calibration of an RGB-D camera couple is an important task in computer vision, especially applications requiring the knowledge of 3D information. Although a variety of algorithms have been proposed, it remains difficult to evaluate existing methods in the literature as different sequences are used. In this paper, we propose a full dataset benchmark, with real and synthetically generated sequences, manually determined ground truth, and evaluation metrics for comparison. Evaluation of three methods using this framework is also provided. The proposed benchmark dataset is available online at http://rfai.lifat.univ-tours.fr/PublicData/ calibrgbd/CalibRGBD.html.
-
Reassembly of Fractured Object Using Fragment Topology
- Author(s): A. Alzaid and S. Dogramadzi
- + Show details - Hide details
-
p.
(NaN)
This work presents our results on reassembly of broken objects using a newly developed fragment topology and feature extraction methodology. The reassembly of broken objects is a common problem in different domains including computeraided bone fracture reduction and reassembly of broken artefacts. The new fragment topology combines information from intact and fractured region boundaries to reduce possible correspondences between the fragments and optimise our iterative matching process. Experiments performed on different multifragment objects show that the proposed topology can be effectively applied, completing the process in a small number of iterations and with average alignment error 0.12mm.
-
Moving Object Detection for Video Satellite Based on Transfer Learning Deep Convolutional Neural Networks
- Author(s): Zhenguo Yan ; Xin Song ; Hanyang Zhong ; Fengqi Jiang
- + Show details - Hide details
-
p.
(NaN)
With the continuous development of the video satellite, its applications in the military and civilian fields are becoming more and more widespread. In recent years, video satellite has become increasingly important to perform real-time detection of moving object, but the influence of various imaging factors usually causes changes in the object features, which greatly increases the difficulty of object detection. Our work aims at the problems of object detection for video satellite, and we apply deep convolutional neural networks (CNNs) to solve these problems for video satellite object detection. In this paper, a deep regression-based CNN object detection method combined with transfer learning is proposed, which can detect the moving object and identify its category in real time. The experiments have shown that the method effectively improves the speed and precision for video satellite object detection under limited samples, which has good generalization on different resolution video images and outperforms other stateof-the-art detection methods. Our method achieves F1-Score of 90.50% in the test set and the average detection time of each image block is about 0.025s, which has lower time overhead and better robustness to rotation and illumination changes of the object.
