Mansour Zuair

Deep Learning Approach for Car Detection in UAV Imagery

This paper presents an automatic solution to the problem of detecting and counting cars in unmanned aerial vehicle (UAV) images. This is a challenging task given the very high spatial resolution of UAV images (on the order of a few centimetres) and the extremely high level of detail, which require suitable automatic analysis methods. Our proposed method begins by segmenting the input image into small homogeneous regions, which can be used as candidate locations for car detection. Next, a window is extracted around each region, and deep learning is used to mine highly descriptive features from these windows. We use a deep convolutional neural network (CNN) system that is already pre-trained on huge auxiliary data as a feature extraction tool, combined with a linear support vector machine (SVM) classifier to classify regions into “car” and “no-car” classes. The final step is devoted to a fine-tuning procedure which performs morphological dilation to smooth the detected regions and fill any holes. In addition, small isolated regions are analysed further using a few sliding rectangular windows to locate cars more accurately and remove false positives. To evaluate our method, experiments were conducted on a challenging set of real UAV images acquired over an urban area. The experimental results have proven that the proposed method outperforms the state-of-the-art methods, both in terms of accuracy and computational time.

A facial expression recognition system using robust face features from depth videos and deep learning

Abstract This work proposes a depth camera-based robust facial expression recognition (FER) system that can be adopted for better human machine interaction. Although video-based facial expression analysis has been focused on by many researchers, there are still various problems to be solved in this regard such as noise due to illumination variations over time. Depth video data in the helps to make an FER system person-independent as pixel values in depth images are distributed based on distances from a depth camera. Besides, depth images should resolve some privacy issues as real identity of a user can be hidden. The accuracy of an FER system is much dependent on the extraction of robust features. Here, we propose a novel method to extract salient features from depth faces that are further combined with deep learning for efficient training and recognition. Eight directional strengths are obtained for each pixel in a depth image where signs of some top strengths are arranged to represent unique as well as robust face features, which can be denoted as Modified Local Directional Patterns (MLDP). The MLDP features are further processed by Generalized Discriminant Analysis (GDA) for better face feature extraction. GDA is an efficient tool that helps distinguishing MLDP features of different facial expressions by clustering the features from the same expression as close as possible and separating the features from different expressions as much as possible in a non-linear space. Then, MLDP-GDA features are applied with Deep Belief Network (DBN) for training different facial expressions. Finally, the trained DBN is used to recognize facial expressions in a depth video for testing. The proposed approach was compared with other traditional approaches in a standalone system where the proposed one showed its superiority by achieving mean recognition rate of 96.25% where the other approaches could make 91.67% at the best. The deep learning-based training and recognition of the facial expression features can also be undertaken with cloud computing to support many users and make the system faster than a standalone system.

Classification of AAMI heartbeat classes with an interactive ELM ensemble learning approach

Abstract In recent years, the recommendations of the Association for the Advancement of Medical Instrumentation (AAMI) for class labeling and results presentation are closely followed as a possible solution for standardization. Regardless of the class normalization, this standard basically recommends for performance evaluation to adopt inter-patient scenarios, which renders the classification task very challenging due to the strong variability of ECG signals. To deal with this issue, we propose in this paper a novel interactive ensemble learning approach based on the extreme learning machine (ELM) classifier and the induced ordered weighted averaging (IOWA) operators. While ELM is adopted for ensemble generation the IOWA operators are used for aggregating the obtained predictions in a nonlinear way. During the iterative learning process, the approach allows the expert to label the most relevant and uncertain ECG heart beats in the data under analysis and then adds them to the original training set for retraining. The experimental results obtained on the widely used MIT-BIH arrhythmia database show that the proposed approach significantly outperforms state-of-the-art methods after labeling on average 100 ECG beats per record. In addition, the results obtained on four other ECG databases starting with the same initial training set from MIT-BIH confirm its promising generalization capability.

Enhancement of mobile robot localization using extended Kalman filter

In this article, we introduce a localization system to reduce the accumulation of errors existing in the dead-reckoning method of mobile robot localization. Dead-reckoning depends on the information that comes from the encoders. Many factors, such as wheel slippage, surface roughness, and mechanical tolerances, affect the accuracy of dead-reckoning. Therefore, an accumulation of errors exists in the dead-reckoning method. In this article, we propose a new localization system to enhance the localization operation of the mobile robots. The proposed localization system uses the extended Kalman filter combined with infrared sensors in order to solve the problems of dead-reckoning. The proposed system executes the extended Kalman filter cycle, using the walls in the working environment as references (landmarks), to correct errors in the robot’s position (positional uncertainty). The accuracy and robustness of the proposed method are evaluated in the experiment results’ section.

Multi-sensors multi-baseline mapping system for mobile robot using stereovision camera and laser-range device

Countless applications today are using mobile robots, including autonomous navigation, security patrolling, housework, search-and-rescue operations, material handling, manufacturing, and automated transportation systems. Regardless of the application, a mobile robot must use a robust autonomous navigation system. Autonomous navigation remains one of the primary challenges in the mobile-robot industry; many control algorithms and techniques have been recently developed that aim to overcome this challenge. Among autonomous navigation methods, vision-based systems have been growing in recent years due to rapid gains in computational power and the reliability of visual sensors. The primary focus of research into vision-based navigation is to allow a mobile robot to navigate in an unstructured environment without collision. In recent years, several researchers have looked at methods for setting up autonomous mobile robots for navigational tasks. Among these methods, stereovision-based navigation is a promising approach for reliable and efficient navigation. In this article, we create and develop a novel mapping system for a robust autonomous navigation system. The main contribution of this article is the fuse of the multi-baseline stereovision (narrow and wide baselines) and laser-range reading data to enhance the accuracy of the point cloud, to reduce the ambiguity of correspondence matching, and to extend the field of view of the proposed mapping system to 180°. Another contribution is the pruning the region of interest of the three-dimensional point clouds to reduce the computational burden involved in the stereo process. Therefore, we called the proposed system multi-sensors multi-baseline mapping system. The experimental results illustrate the robustness and accuracy of the proposed system.

Tile-Based Semisupervised Classification of Large-Scale VHR Remote Sensing Images

This paper deals with the problem of the classification of large-scale very high-resolution (VHR) remote sensing (RS) images in a semisupervised scenario, where we have a limited training set (less than ten training samples per class). Typical pixel-based classification methods are unfeasible for large-scale VHR images. Thus, as a practical and efficient solution, we propose to subdivide the large image into a grid of tiles and then classify the tiles instead of classifying pixels. Our proposed method uses the power of a pretrained convolutional neural network (CNN) to first extract descriptive features from each tile. Next, a neural network classifier (composed of 2 fully connected layers) is trained in a semisupervised fashion and used to classify all remaining tiles in the image. This basically presents a coarse classification of the image, which is sufficient for many RS application. The second contribution deals with the employment of the semisupervised learning to improve the classification accuracy. We present a novel semisupervised approach which exploits both the spectral and spatial relationships embedded in the remaining unlabelled tiles. In particular, we embed a spectral graph Laplacian in the hidden layer of the neural network. In addition, we apply regularization of the output labels using a spatial graph Laplacian and the random Walker algorithm. Experimental results obtained by testing the method on two large-scale images acquired by the IKONOS2 sensor reveal promising capabilities of this method in terms of classification accuracy even with less than ten training samples per class.

Domain Adaptation Network for Cross-Scene Classification

In this paper, we present a domain adaptation network to deal with classification scenarios subjected to the data shift problem (i.e., labeled and unlabeled images acquired with different sensors and over completely different geographical areas). We rely on the power of pretrained convolutional neural networks (CNNs) to generate an initial feature representation of the labeled and unlabeled images under analysis, referred as source and target domains, respectively. Then we feed the resulting features to an extra network placed on the top of the pretrained CNN for further learning. During the fine-tuning phase, we learn the weights of this network by jointly minimizing three regularization terms, which are: 1) the cross-entropy error on the labeled source data; 2) the maximum mean discrepancy between the source and target data distributions; and 3) the geometrical structure of the target data. Furthermore, to obtain robust hidden representations we propose a mini-batch gradient-based optimization method with a dynamic sample size for the local alignment of the source and target distributions. To validate the method, in the experiments we use the University of California Merced data set and a new multisensor data set acquired over several regions of the Kingdom of Saudi Arabia. The experiments show that: 1) pretrained CNNs offer an interesting solution for image classification compared to state-of-the-art methods; 2) their performances can be degraded when dealing with data sets subjected to the data shift problem; and 3) how the proposed approach represents a promising solution for effectively handling this issue.

Ear recognition via sparse coding of local features

Abstract. An efficient scheme for human ear recognition is presented. This scheme comprises three main phases. First, the ear image is decomposed into a pyramid of progressively downgraded images, which allows the local patterns of the ear to be captured. Second, histograms of local features are extracted from each image in the pyramid and then concatenated to shape one single descriptor of the image. Third, the procedure is finalized by using decision making based on sparse coding. Experiments conducted on two datasets, composed of 125 and 221 subjects, respectively, have demonstrated the efficiency of the proposed strategy as compared to various existing methods. For instance, scores of 96.27% and 96.93% have been obtained for the datasets, respectively.

False Positive Reduction in Breast Mass Detection Using the Fusion of Texture and Gradient Orientation Features

The presence of masses in mammograms is among the main indicators of breast cancer and their diagnosis is a challenging task. The one problem of Computer aided diagnosis (CAD) systems developed to assist radiologists in detecting masses is high false positive rate i.e. normal breast tissues are detected as masses. This problem can be reduced if localised texture and gradient orientation patterns in suspicious Regions Of Interest (ROIs) are captured in a robust way. Discriminative Robust Local Binary Pattern (DRLBP) and Discriminative Robust Local Ternary Pattern (DRLTP) are among the state-of-the-art best texture descriptors whereas Histogram of Oriented Gradient (HOG) is one of the best descriptor for gradient orientation patterns. To capture the discriminative micro-patterns existing in ROIs, we propose localised DRLBP-HOG and DRLTP-HOG descriptors by fusing DRLBP, DRLTP and HOG for the description of ROIs; the localisation is archived by dividing each ROI into a number of blocks (sub-images). Support Vector Machine (SVM) is used to classify mass or normal ROIs. The evaluation on DDSM, a benchmark mammograms database, revealed that localised DRLBP-HOG with 9 (3\(\times \)3) blocks forms the best representation and yields an accuracy of 99.80±0.62(ACC±STD) outperforming the state-of-the-art methods.

PH-BRINT: Pooled Homomorphic Binary Rotation Invariant and Noise Tolerant Representation for Face Recognition Under Illumination Variations

Face recognition under varying illumination conditions is a challenging problem. We propose a simple and effective multiresolution approach Pooled Homomorphic Binary Rotation Invariant and Noise Tolerant (PH-BRINT) for face recognition under varying illumination conditions. First, to reduce the effect of illumination, wavelet transform based homomorphic filter is used. Then Binary Rotation Invariant and Noise Tolerant (BRINT) operators with three different scales are employed to extract multiscale local rotation invariant and illumination insensitive texture features. Finally, the discriminative information from the three scales is pooled using MAX pooling operator and localized gradient information is computed by dividing the pooled image into blocks and calculating the gradient magnitude and direction of each block. The PH-BRINT technique has been tested on a challenging face database Extended Yale B, which was captured under varying illumination conditions. The system using minimum distance classifier with L1-norm achieved an average accuracy of 86.91%, which is comparable with the state-of-the-art best illumination-invariant face recognition techniques.