Eanes Torres Pereira

Integral Local Binary Patterns: A Novel Approach Suitable for Texture-Based Object Detection Tasks

This work is concerned with the proposition and empirical evaluation of a new feature extraction approach that combines two existing image descriptors, Integral Histograms and Local Binary Patterns (LBP), to achieve a representation that exhibits relevant properties to object detection tasks (such as face detection): fast constant time processing, rotation, and scale invariance. This novel approach is called the Integral Local Binary Patterns (INTLBP), which is based on an existing method for calculating Integral Histograms from LBP images. This paper empirically demonstrates the properties of INTLBP in a scenario of texture extraction for face/non-face classification. Experiments have shown that the new representation added robustness to scale variations in the test images - the proposed approach achieved a mean classification rate 92% higher than the standard Rotation Invariant LBP approach, when testing over images with scales different from the ones used for training. Moreover, the INTLBP dramatically reduced the required processing time when searching patterns in a face detection task.

Investigation of local and global features for face detection

This work is concerned with the empirical evaluation of a set of local and global features under the context of frontal (including semi-profile) and full profile face classification. Integral LBP, Integral Histograms, PCA and Optimized Face Ratios features have been evaluated using SVM classifiers. A data set of about 14,000 face and 300,000 non face images has been used in the experiments. Face images were obtained from well known public face research databases, such as BioID, Color Feret, CMU PIE, among others. The PCA-SVM classifier presented best overall results for both frontal and full profile faces whereas the classifier based on Face Ratios presented the lowest classification rates. A weighted combination of all classifiers yielded high True Positive (TPR) and True Negative (TNR) rates: 91.7% and 100%, respectively, for the frontal face experiments; 99.59% and 99.62%, respectively, for the profile face experiments. These results indicate that the evaluated features are very suitable to the problem of face detection and that a simple classifier combination improves individual classifiers performance.

Local Binary Patterns Applied to Breast Cancer Classification in Mammographies

Among all cancer types, breast cancer is the one with the second highest incidence rate for women. Mammography is the most used method for breast cancer detection, as it reveals abnormalities such as masses, calcifications, asymmetries and architectural distortions. In this paper, we propose a classification method for breast cancer that has been tested for six different cancer types: CALC, CIRC, SPIC, MISC, ARCH, ASYM. The proposed approach is composed of a SVM classifier trained with LBP features. The MIAS image database was used in the experiments and ROC curves were generated. To the best of our knowledge, our approach is the first to handle those six different cancer types using the same technique. One important result of the proposed approach is that it was tested over six different breast cancer types proving to be generic enough to obtain high classification results in all cases.

An Approach for Multi-pose Face Detection Exploring Invariance by Training

In this paper, a rotation invariant approach for face detection is proposed. The approach consists of training specific Haar cascades for ranges of in-plane face orientations, varying from coarse to fine. As the Haar features are not robust enough to cope with high in-plane rotations over many different images, they are trained only until an accented decay in precision is evident. When that happens, the range of orientations is divided up into sub-ranges, and this procedure continues until a predefined rotation range is reached. The effectiveness of the approach is evaluated on a face detection problem considering two well-known data sets: CMU-MIT[1] and FDDB[2]. When tested using CMU-MIT dataset, the proposed approach achieved accuracies higher than the traditional methods such as the ones proposed by Viola and Jones[3] and Rowley et al.[1]. The proposed approach has also achieved a large area under the ROC curve and true positive rates that were higher than the rates of all the published methods tested over the FDDB dataset.

Automatic and semi-automatic annotation of people in photography using shared events

This article proposes an automatic and semi-automatic annotation technique for people in photos using the shared event concept, which consists of many photos captured by different devices of people who attended the same event. The technique uses an algorithm to group photos into personal events and then verifies which of these events are shared. The automatic annotation of people uses techniques of facial recognition and detection, while the semi-automatic annotation uses a pondered sum of estimators based on contextual information and picture content. Experiments showed that using the shared event concept increases the hit rate of automatic and semi-automatic annotations of people in the utilized photo collection.

Improving Face Detection Performance by Skin Detection Post-Processing

Face detection is already incorporated in many biometrics and surveillance applications. Therefore, the reduction of false detections is a priority in those systems. However, face detection is still challenging. Many factors, such as pose variation and complex backgrounds, contribute to false detections. Besides, the fidelity of a true detection, measured by precision rate, is a concern in content-based information retrieval. Following those issues, combinations of methods are developed focusing on balancing the trade-off between hit-rate and miss-rate. In this paper, we present an approach that improves face detection based on a post-processing of skin features. Our method enhanced the performance of weak detectors using a straightforward and low complex skin percentage threshold constraint. Furthermore, we also present a statistical analysis comparing our approach and two face detectors, under two different conditions for skin detection training, using a robust dataset for testing. The experimental results showed a significant drop in the number of false positives, reducing in 53%, while the precision rate was elevated in almost 5% when the Viola-Jones approach was used as face detector.

A robust music genre classification approach for global and regional music datasets evaluation

This paper deals with two problems: (1) the selection of a set of music features in order to achieve high genre classification accuracies; (2) the absence of a representative music dataset of regional Brazilian music. In this paper, we propose a set of features to classify genres of music. The features proposed were obtained by a methodical selection of important features used in the literature of Music Information Retrieval (MIR) and Music Emotion Recognition (MER). Besides, we propose a new music dataset called BMD (Brazilian Music Dataset)1, containing 120 songs labeled in 7 musical genres: FoFFó, Rock, Repente, MPB(Música Popular Brasileira — Brazilian Popular Music), Brega, Sertanejo and Disco. An important characteristic of this new dataset compared with others, is the presence of three popular genres in Brazil Northeast region: Repente, Brega and a characteristic genre similar to MPB, which we also call as MPB. We evaluated our proposed features on both datasets: GTZAN and BMD. The proposed approach achieved average accuracy (after 30 runs of 5-fold-cross-validations) of 79.7% for GTZAN and 86.11% for the BMD. Another important contribution of this work is random repetition of cross-validation executions. Most of the papers performs only a single n-fold cross-validation. We criticize that practice and propose, at least, 30 random executions to compute the average accuracy.

The role of data balancing for emotion classification using EEG signals

In this paper, we demonstrate the role of data balancing in experimental evaluation of emotion classification systems based on electroencephalogram (EEG) signals. ADASYN method was employed to create a balanced version of the DEAP EEG dataset. Experiments considered Support Vector Machine classifiers trained with HOC and PSD features to predict valence and arousal affective dimensions. Using signals from only four channels (Fp1, Fp2, F3 and F4) we obtained, after balancing, accuracies of 98% (valence) and 99% (arousal) for subject dependent experiments with three classes, and 85% (valence) and 87% (arousal) for two-class classification. However, accuracies for subject independent experiments were lower than the ones obtained using imbalanced datasets. We obtained accuracies of 52% (valence) and of 49% (arousal) for two classes, and accuracies of 36% (valence) and of 31% (arousal) for three classes. To explain the low accuracies in subject independent experiments, we present arguments and empirical evidence using correlations between the percentage of samples for each class and the accuracies obtained by approaches which did not use balanced datasets.

Guiding a bottom-up visual attention mechanism to locate specific image regions using a distributed genetic optimization

The purpose of this paper is to present an approach to locate specific regions in images. The novelty of the approach is the combination of a weighted bottom-up visual attention mechanism with a genetic algorithm optimization running on a computational grid. The visual attention mechanism is based on the model proposed by Itti and Koch [1]. A saliency map indicates the most interesting points in an image using a number of intermediate low level features, which are detected at different scales and orientations. Using the saliency map weights as parameters, the optimization problem is to minimize the number of most salient points needed to locate a set of reference image regions, previously (and manually) labeled as being interesting. Both an objective and subjective evaluation have demonstrated that the proposed approach is more effective when compared to a fixed weight attention mechanism.