Brazil

In this paper, we propose a method for face recognition
by applying data augmentation and transfer learning in pretrained Convolutional Neural Networks (CNNs). Our main focus
is to analyze the power of data augmentation for face recognition
systems with CNN transfer learning. We extracted features from
the images and trained a KNN classifier using the VGG-Face
CNN.

IWSSIP 2020

Data Augmentation for Face Recognition with CNN Transfer Learning

face recognition

data augmentation

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

technical paper

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

Around the world, there are many people with disabilities; it is estimated that 39 million people are blind and 246 million have limited vision, giving a total of 285 million visually impaired people. The use of information and communication technologies can help disabled people to achieve greater independence, quality of life and inclusion in social activities by increasing, maintaining or improving their functional capacities. In this context, this paper presents an automatic methodology for identifying banknotes that can be widely used by people with visual impairment. For this, we evaluated a set of four point of interest detectors, two descriptors, seven ways of generating the image signature, and six classification methodologies, which can be used as a basis for the development of applications for the identification of banknotes. Experiments performed on US Dollar (USD), Euro (EUR) and Brazilian Real Banknotes (BRL) obtained rates of accuracy of 99.78%, 99.12%, and 96.92%, respectively.

Banknote Identification Methodology for Visually Impaired People

Glaucoma is identified as one of the main global
causes of visual impairment or blindness. There is no cure or possible corrections in case of visual impairment, so the early diagnosis is essential to delay or prevent its progression. However, the disease is asymptomatic in its early stages. The disease can be detected in routine eye exams like anterior segment optical coherence tomography, which is analyzed by a specialist, but the analysis of many patients demands much time. There are two main types of the disease, open angle and angle closure glaucoma. In this paper, we propose a method for automatic detection of angle closure glaucoma in anterior segment optical coherence tomography images, based on transfer learning and multilevel convolutional neural networks to extract visual features. In the proposed method, the multilevel architecture models achieve as the best result an AUC of 0.972."

Multilevel CNN for Angle Closure Glaucoma Detection using AS-OCT Images

One of the most used statistical approaches in the
literature for computing texture information is the Gray Level Cooccurrence Matrix (GLCM). The use of GLCM and textures descriptors calculated from there, allows, among other applications, the detection of the main direction of the elements of texture. This is possible essentially because the GLCM uses different angles to compute the relationship between the gray
tones in the image. An alternative to computing this relationship is the Sum of Difference Histograms (SDH). In contrast to GLCM, which is a bidimensional structure, SDHs are onedimensional. Once there is a relationship between the data stored in both structures, the behavior of the descriptors is similar, and the SDH have a lower computational cost being, for this, more efficient in a number of applications. In this sense, this work leads to a series of investigations regarding the correspondence between
these two approaches. Moreover, it investigates which texture descriptors, calculated from the SDH, can collaborate in the process of identifying the predominant direction of a texture.

Determining the Direction of Images Using Sum and Difference Histograms and Texture Descriptors

This work proposes the use of Genetic Algorithms (GA) to identify the area of the breast from the background in thermographic breast images. The proposed method uses color information, a fitness function based on cardioids, and GA. This is the first work in the literature to propose a Region of Interest (ROI) extraction based on GA and cariods. ROI extraction can improve the accuracy of cancer detection and assist with the standardization of acquisition protocols. The method is able to successfully separate the breast region in 52 out of 58 images, while being fully automatic, and not requiring manual selection of seed points.

ROI Extraction in Thermographic Breast Images Using Genetic Algorithms

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.

Hybrid Capsule Network Architecture Estimation for Melanoma Detection

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.

Premium content

Downloads

Next from IWSSIP 2020

Banknote Identification Methodology for Visually Impaired People

Similar lecture

CNN Based Predictor of Face Image Quality