Brazil

Hybrid Capsule Network Architecture Estimation for Melanoma Detection. As uncommon as its incidence is, melanoma is considered to be one of the most aggressive and deadly skin cancers today. Risk factors for the development of the disease are related to exposure to ultraviolet rays and family history. Currently, the diagnosis is made through dermatoscopy, a noninvasive technique that aims to observe the structures of the skin. The diagnosis of melanoma is linked to the professional&#39;s experience in identifying the disease. With the advancement of computing, solutions based on deep learning were suggested to help diagnose the disease. This work has as main objective the adaptation of a capsule neural network for the task of detecting melanoma in medical images. The proposed architecture consists of the VGG16 neural network combined with a capsule neural network. The results obtained by the proposed model fed with the ISIC Archive database proved to be promising when compared to the literature, obtaining an accuracy of 92.6, sensitivity of 90.9 and specificity of 92.8.

IWSSIP 2020

Hybrid Capsule Network Architecture Estimation for Melanoma Detection

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

Hybrid Capsule Network Architecture Estimation for Melanoma Detection. As uncommon as its incidence is, melanoma is considered to be one of the most aggressive and deadly skin cancers today. Risk factors for the development of the disease are related to exposure to ultraviolet rays and family history. Currently, the diagnosis is made through dermatoscopy, a noninvasive technique that aims to observe the structures of the skin. The diagnosis of melanoma is linked to the professional's experience in identifying the disease. With the advancement of computing, solutions based on deep learning were suggested to help diagnose the disease. This work has as main objective the adaptation of a capsule neural network for the task of detecting melanoma in medical images. The proposed architecture consists of the VGG16 neural network combined with a capsule neural network. The results obtained by the proposed model fed with the ISIC Archive database proved to be promising when compared to the literature, obtaining an accuracy of 92.6, sensitivity of 90.9 and specificity of 92.8.

technical paper

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

One of the fundamental dilemmas of mobile robotics is the use of sensory information to locate an agent in geographic space. In this paper, we developed a global relocation system to predict the robot’s position and avoid unforeseen actions from a monocular image, which we named SpaceYNet. We incorporated Inception layers to symmetric layers of down-sampling and up-sampling to solve depth-scene and 6-DoF estimation simultaneously. Also, we compared SpaceYNet to PoseNet - a state of the art in robot pose regression using CNN - in order to evaluate it. The comparison comprised one public dataset and one created in a broad indoor environment. SpaceYNet showed higher accuracy in global percentages when compared to PoseNet.

SpaceYNet: A Novel Approach to Pose and Depth-Scene Regression Simultaneously

Recently, ocular biometrics has been gaining importance in Biometrics due to the poor performance obtained in some cases by biometric systems based on characteristics of the whole face. This paper presents a new method for person authentication based on ocular deep features, which are extracted from the ocular region of the face by using a very deep CNN (Convolutional Neural Network). Another interesting aspect of our method is that, instead of using directly the deep features as input for the authentication system, it uses the difference between the probe and gallery deep features. So, our method adopts a pairwise strategy. A binary support vector machine is trained to determine whether a given difference vector is genuine or impostor. The proposed new method based on such pairwise strategy was evaluated using the ocular left set of the UBIPr dataset and five pre-trained CNN architectures. When using the pre-trained VGG-Face the proposed method obtained a state-of-the-art result (3.18% of Equal Error Rate).


Ocular Recognition Using Deep Features for Identity Authentication

This work explores and compares the plethora of metrics for the performance evaluation of object-detection algorithms. Average precision (AP), for instance, is a popular metric for evaluating the accuracy of object detectors by estimating the area under the curve (AUC) of the precision × recall relationship. 
Depending on the point interpolation used in the plot, two different AP variants can be defined and, therefore, different results are generated. AP has six additional variants increasing the possibilities of benchmarking. The lack of consensus in different works and AP implementations is a problem faced by the academic and scientific communities. Metric implementations written in different computational languages and platforms are usually distributed with corresponding datasets sharing a given bounding box description. Such projects indeed help the community with evaluation tools, but demand extra work to be adapted for other datasets and bounding-box formats. This work reviews the most used metrics for object detection detaching their differences, applications, and main concepts. It also proposes a standard implementation that can be used as a benchmark among different datasets with minimum adaptation on the annotation files.

A Survey on Performance Metrics for Object-Detection Algorithms

Deep learning-based classifiers for object recognition and classification have been used in the domain of plant disease detection, particularly lesions from leaf images. In such domains, as expected, deep neural networks perform better using balanced data sets than imbalanced ones, as they exhibit some inductive bias favoring balanced data from each class. However, data sets for plant disease detection are often imbalanced due to the rarity of disease lesions in real-world settings. While deep generative approaches such as generative adversarial networks (GANs) have been established as an effective means of augmenting high-dimensional image data, the literature lacks a detailed study of the effectiveness of GAN-based models on a plant disease detection task, compared to sampling-based approaches traditionally used to reduce the skewness of the data. In this paper, we comparatively evaluate an image classifier based on a dense convolutional neural network (CNN), trained using a GAN, versus the same CNN model used in tandem with undersampling, oversampling, and an adaptation of the Synthetic Minority Over-sampling Technique (SMOTE). The GAN-based approach is shown to attain significantly higher recall and hence F-measure and ROC AUC against each of these.

Addressing Class Imbalance in Image-Based Plant Disease Detection: Deep Generative vs. Sampling-Based Approaches

Breast cancer is one of the most common diseasesin women in the world. There are various imaging techniquesemployed in the diagnosis. The histological image analysis sup-ported by computational systems has proved to be quite effectivein diagnosing the disease. In this paper, we present an approachto quantify and classify tissue samples of the breast based on features extracted from the intensity histogram, co-occurrencematrix and the Shannon, Renyi, Tsallis and Kapoor entropies. The attribute set was employed to obtain the feature vectors which were evaluated as inputs to the random forest and sequential minimal optimization algorithms with the 10-fold cross-validation technique. In this study, we investigated the proposed approach with images obtained in four levels of magnification of the publicly available Breast Cancer Histopathological Database. In the feature selection stage, we investigated the correlation-Based feature selection, ReliefF, information gain, gain ratio, one-R and symmetrical uncertainty algorithms for evaluating the performance of theproposed approach. The proposed approach achieved significant results of AUC and accuracy for all cases analyzed. The proposed approach obtained 0.997 for AUC and 97.6% for the accuracy metric. These results are considered relevant and this approachis useful as an automated protocol for the diagnosis of breast histological tissue.

Analysis of Features for Breast Cancer Recognition in Different Magnifications of Histopathological Images

Borehole stability is one of the main factors influencing the success of an oil well drilling process. Fractures and beddings (sedimentary layers) are important stability features commonly identified in borehole imaging tools. Their visual presentation in the 2D Ultrasonic Borehole Imager (UBI) consists of sinusoidal curves, in which automatic detection is not trivial. Several previous works proposed different solutions, most of them based on the Hough transform (HT). Nevertheless, HT based solutions have a substantial computational cost. In this work, we propose an innovative approach to identify the harmonic curves related to fractures and sedimentary layers in borehole images based in the Iterative Local Search heuristic, reducing the computational cost of the task significantly.

Automatic Sinusoidal Curves Detection in Borehole Images Using the Iterated Local Search Algorithm

This paper presents an approach for evaluating CAD systems using concepts of Content-Based Image Retrieval and graphical oracles. After implementing feature descriptors and selecting three similarity functions, two metrics are proposed to measure the efficiency of CAD systems. A case study was conducted considering three simulated CAD systems to detect masses and calcifications in mammographic images. The results indicated that the our approach is as robust as traditional metrics with respect to performance evaluation. However, our approach is more flexible than traditional metrics because evaluators can choose the more adequate features to assess a particular CAD system.

Graphical Oracles to Assess Computer-Aided Diagnosis Systems: A Case Study in Mammogram Masses and Calcifications Detection

Active Shape Models (ASM) are widely explored for facial landmarks detection. The task is commonly approached in frontal perspective where target features such as eyes, mouth, and nose lie within the face’s contour. Conversely, few application proposals are aimed to find facial landmarks in a profile view. This paper addresses an issue with that technique applied to locate facial landmarks lying on the profile contour. Those boundary points present an extra challenge for ASM’s searching method when applied in images with background lighting variations. An ASM trained with profile images captured with different brightness backgrounds may lead the method to model texture profiles in segregated clusters. In those cases, it is inaccurate to assume that sampled texture profiles follow a Gaussian distribution, which is essential to estimate new landmarks locations with the Mahalanobis distance. We propose an alternative to improve ASM’s robustness in such a scenario by utilizing K- Means to identify those clusters and create specific texture models for each one. Then, during the searching stage, new estimates are compared with each model individually, assuming the texture profiles follow a normal distribution inside their corresponding clusters. We trained and tested our improved ASM with a private database of 670 face images in profile view captured in a controlled environment. We report an average error reduction of 46% for landmarks location compared to the classical method. We also demonstrate that, regarding this issue, our proposal drives ASM’s searching method to the point of convergence, whereas the original method does not.

Improving Active Shape Models Robustness Towards Locating Facial Landmarks In Profile Contour

In this paper, we proposed and evaluated the adoption of a capsule network-based (CapsNet-based) model rather than the convolutional neural network-based (CNN-based) models which are predominant in crowd counting tasks. The aim is to join the task of generating a high-quality density map from a single image along with producing a more precise estimate of the number of people. CapsNet-based model has a strong capacity of representation and a powerful dynamic routing mechanism that could address the drawback of a limited number of training samples. The replacement of the scalar values of CNN by vectors when using a CapsNet allows learning more discriminative features, which contributes to generating high-quality density maps, and thus a more precise number of individuals in crowd scenes. Experimental results show that our proposal presents competitive results concerning state-of-the-art, but with a 59.2% reduction in the number of parameters.

Dense Crowd Counting with Capsule Networks

Acute leukemia is a cancer-related to a bone marrow abnormality. It is more common in children and young adults. This type of leukemia generates unusual cell growth in a short period, requiring a quick start of treatment. Acute Lymphoid Leukemia (ALL) and Acute Myeloid Leukemia (AML) are the main responsible for deaths caused by this cancer. The classification of these two leukemia types on blood slide images is a vital process of and automatic system that can assist doctors in the selection of appropriate treatment. This work presents a convolutional neural networks (CNNs) architecture capable of differentiating blood slides with ALL, AML and Healthy Blood Slides (HBS). The experiments were performed using 16 datasets with 2,415 images, and the accuracy of 97.18% and a precision of 97.23% were achieved. The proposed model results were compared with the results obtained by the state of the art methods, including also based on CNNs.

Premium content

Downloads

Next from IWSSIP 2020

SpaceYNet: A Novel Approach to Pose and Depth-Scene Regression Simultaneously

Similar lecture

A Comparison of Neural Network Approaches for Melanoma Classification