Brazil

This work presents a technique of working together to accumulate learning between various artificial neural networks, the technique has been successfully adapted to a convolutional network and has shown promising results in the context of traffic light classification.

IWSSIP 2020

Optimizing CNN&#39;s using Cumulative Learning with Auxiliary Networks for Traffic Light Recognition

multiple self-coordinated neural networks

traffic lights recognition

deep learning

ICPR 2020 is the premier international conference on Pattern Recognition, gathering leading researchers from around the world to present their latest achievements in Pattern Analysis and Recognition, applied Artificial Intelligence and Machine Learning. The topic of the year is " Putting Artificial intelligence to work on Patterns". 
For the first time, ICPR has employed a two-round review process similar to journal submissions, with new papers allowed to be submitted in either the first or second round and papers submitted in the first round and not accepted allowed to be revised and re-submitted for second round review. 

In this edition of the **ICPR 2020** Conference, there are 1407 Oral/Poster presentations organized in five Tracks: 
- Artificial Intelligence, Machine Learning for Pattern Analysis; 
- Biometrics, Human Analysis and Behavior Understanding; Computer Vision, Robotics and Intelligent Systems; 
- Document and Media Analysis; and 
- Image and Signal Processing. 

The **ICPR 2020** Main Conference program also includes 7 Keynotes from prominent world-renowned scientists, 8 Tutorials, 28 Demos, 8 Challenges, 2 Panels, as well as an Industry/Sponsor Exhibition. In addition, 41 Workshops on a range of important current and emerging research topics complete the ICPR 2020 program. 


ICPR 2020

Optimizing CNN's using Cumulative Learning with Auxiliary Networks for Traffic Light Recognition

technical paper

The 27th International Conference on Systems, Signals and Image Processing.

This presentation shows a study about the feasibility of combining sustained vowels for computer-based pathological voice characterization. Although recent work has shown that single sustained vowels enable accurate voice characterization, the literature lacks evidence about combining them. Here we propose a pipeline for pathological voice characterization and classification that encompasses a feature extraction method based on the Haar wavelet. We also evaluate the benefits of our pipeline when using an input signal that comprises three different sustained vowels instead of single ones. We conducted experiments on samples of sustained vowels /a/, /i/, and /u/ from two public datasets, exploring the wavelet decomposition levels in the range of 4 to 18.

Combined Sustained Vowels Improve the Performance of the Haar Wavelet for Pathological Voice Characterization

The skin is an essential system for the human body, also the most susceptible to diseases. The diagnosis of these diseases is made mainly through dermatoscopy images, a noninvasive procedure. This paper presents a semi-automatic method that can assist the physician in monitoring the skin lesion’s evolution. Our approach applies the SLIC0 algorithm for superpixels generation, and texture features extraction. Subsequently, the algorithm Seeded Fuzzy C-means clusters the superpixels based on marked regions made by a specialist. Finally, the images undergo a post-processing step, eliminating noise and smoothing edges. We used 1380 images from four public datasets: the PH2 and DermIs to settle the method parameters, and ISIC 2016 and ISIC 2017 to the performance evaluation. Our tests show that by marking 1% of the superpixels, our method achieves an accuracy of 96.16% in ISIC 2016 and 95.34% in ISIC 2017. According to the literature, the segmentation was considered excellent in 93.49% of the images in ISIC 2016 and on 79.44% of ISIC 2017 images.

A Skin Lesion Semi-Supervised Segmentation Method

The chest radiography is one of the most accessible radiological examinations for diagnosis of lung and heart diseases. Deep learning techniques have been increasingly used to provide more accurate detection of thorax lesions on Chest X-Ray (CXR) images. However, we observe that we can use the complementarity of dual asymmetric deep convolutional neural networks (DCNNs) to improve the ability of CXR image classification compared to the single network. In this paper, we propose a novel dual lesion attention network named DuaLAnet for the classification of 14 thorax diseases on chest radiography. The DuaLAnet consists of two asymmetric attention networks, DenseNet-169 and ResNet-152, to integrate the advantages into a wider architecture, thus extracting more discriminative features of different abnormalities from the raw CXRs. Moreover, a training strategy is designed to integrate the loss contribution of the involved classifiers into a unified loss.

DuaLAnet: Dual Lesion Attention Network for Thoracic Disease Classification in Chest X-Rays

In prostate analysis clinical routine, manual and semi-manual techniques are used to interpret MRI scans, which in addition to being subjective, are time-consuming and operator-dependent. Prostate segmentation is necessary for the diagnosis of a possible tumor, biopsy, staging, monitoring and treatment. Thus, an automatic segmentation can be crucial for fast and reliable diagnosis, but is challenging to accomplish due to the varying sizes, shapes and also unclear prostate boundaries that are blended with surrounding tissues. Thereupon, this work proposes an automatic method for prostate segmentation on 3D MRI scans based on superpixels, taxonomic indexes, probabilistic atlas and the XGBoost algorithm. In addition, the proposed method has been evaluated on the Prostate 3T and PROMISE12 databases, presenting a dice similarity coefficient of 84.03%, volumetric similarity of 95.42%, recall of 88.11%, specificity of 90.82%, and an accuracy of 90.06\%. Experimental results demonstrate the high potential of the proposed method comparable to those previously published.

Taxonomic Indexes for Automatic Prostate Segmentation on 3D MRI Scans Using Superpixels and Probabilistic Atlas

Precision agriculture (PA) has greatly benefited from new technologies over the years. The use of multispectral and hyperspectral sensors coupled to Unmanned Aerial Vehicles (UAV) has enabled farmers to monitor crops, improve the use of resources and reduce costs. Despite being widely used, multispectral images present a natural misalignment among the various spectra due to the use of different sensors. The variation of the analyzed spectrum also leads to a loss of characteristics among the bands which hinders the feature detection process among them, which makes the alignment process complex. In this paper, is proposed a new framework for the band co-registration process based on the premise that natural misalignment is an attribute of the camera, so it does not change during the acquisition process. The results were compared with the ground-truth generated by a specialist and with other methods present in the literature. The proposed framework had an average back-projection (BP) error of $0.425$ pixels, this result being $335\%$ better than the evaluated frameworks.

UAV-Multispectral Sensed Data Band Co-Registration Framework

One important strength of adaptive filtering algo-rithms is their ability to learn from their interactions with theenvironment. 
Such a flexibility comes with a price, since theirperformance strongly depends on adjustable parameters. 
In orderto provide to the designer guidelines and guaranties for bothperformance and stability, the open literature often relies onstochastic models that predict the learning capabilities of differentalgorithms.
Unfortunately, usually some statistical hypotheses are assumedto make the mathematics tractable. 
The most critical of themis the independence assumption, which states that the adaptivecoefficients are statistically independent from the excitation data.
Such an assumption hampers the accuracy of the theoreticalpredictions in the case of large step sizes, which is a commonchoice in practice. 
Fortunately, a technique that avoids theindependence assumption, respectively exact expectation analysis,has experienced some advances. 
In this paper, it is demonstratedthat such recent advances can be used to address a complexconfiguration of the least mean squares algorithm, where theadditive noise is colored and the ideal plant the adaptive schemeintends to estimate is nonlinear.

Recent Advances in Overcoming the Independence Assumption in Adaptive Filtering Analyses

The development of efficient tools for conducting measures to characterize the degeneration of brain structures is useful to facilitate early detection and monitoring of the evolution of neurodegenerative diseases such as dementia and Alzheimer's. It is well-known that, with the evolution of these diseases, certain brain structures especially the corpus callosum (CC), progressively diminish in dimension. This work considers using Magnetic Resonance Imaging (MRI) and delimiting the tissues of interest to calculate the volume of this structure. This delimitation or segmentation if done using computational techniques offers reproducibility, in addition to increasing the speed and more precision in calculating the volume variation. Therefore, a study of techniques of automatic segmentation of the CC is presented for the volumetric evaluation of MRI images weighted in T1, in the DICOM format, where the calculated volume is independent of the equipment, allowing the estimation of the progression of the loss of volume of such a structure.

Comparing the Degeneration of the Corpus Callosum by Magnetic Resonance Images

This paper presents the interim results of the ongoing CORELA Erasmus+ project. The aim of the CORELA project is to increase the effectiveness of electrotechnical engineering education and higher the attraction of the specific engineering field. The purpose of the project is to develop a common platform, which can be used in the education process at different levels of higher school education. The developed platform has different operation modes and can be used as a tool for analytical calculation and theoretical exercise. The platform has integrated different models of electronic components, which can be used on a simulation of the real scenarios. The most important future of the platform is real-time operation mode with the engagement of real measurement types of equipment and electronic components. With the capability of internet connection and exchange the data with the Massive Open Online Course (MOOC) platform, the system can be used for Remote Virtual Laboratories (VRL) and collaborative learning. A common MOOC and CORELA platform enable pupils to interact with each other in a manner to share their suggestion and experiences to solve exercise efficiently.

CORELA Collaborative Learning Environment for Electrical Engineering Education

Image Registration is first step in analysis of the same point of the patient over some time. This work proposes a new approach to define the same anatomical point in a set of images without the intensity of the content or the constant intensity hypothesis. It considers Möbius transformation of matched points having affine and projective transformations as particular cases. The experiments consider breast infrared dynamic examination of 23 patients, each one with 20 images.
They are registered in two diverse ways in a total of 874 registrations. Results are evaluated qualitatively and quantitatively. The used examinations are public available and the achieved results are even compared with other well know approach resulting 1748 registration on real images. The proposed approach improves the image matching in 92%, 92%, 81%, 85%, and 81% according to Dice, Jaccard, Superposition of Reference, False Negative and False Positive respectively.

A Non Linear Registration Without the Use of the Brightness Constancy Hypothesis

Glaucoma is a group of eye diseases characterized by progressive degeneration of the optic nerve, which can lead the patient to blindness, however it can be prevented through diagnosis and treatment of the disease. It is the leading cause of irreversible blindness worldwide and is expected to affect 76 million by 2020. The cup-to-disc ratio technique is a structural examination of the optic nerve used to glaucoma diagnose, and requires an accurate segmentation of the optic disc and cup. Thus, this work proposes an automated method for cup segmentation in fundus images of DRISHTI-GS dataset using a combination of the green channel of RGB image and the provided optic disc mask as input to a modified U-Net convolutional neural network. The results achieved were promising, yielding a mean dice coefficient of 94%, much better than state-of-the-art algorithms. The method could be used for glaucoma diagnosis.

Premium content

Downloads

Next from IWSSIP 2020

Combined Sustained Vowels Improve the Performance of the Haar Wavelet for Pathological Voice Characterization

Similar lecture

Transfer Learning with Graph Neural Networks for Short-Term Highway Traffic Forecasting