Mehrez Abdellaoui

Cereal varieties classification using wavelet techniques combined to multi-layer neural networks

This paper presents a new classification method of the various cereal grains varieties. The first phase consists in generating primitives using the wavelet techniques. These primitives are tested by a statistical study and validation tests to extract the deterministic parameters. The second part consists in developing a neuronal classifier designed using the multilayer neural networks to classify the three grain classes (hard wheat, tender wheat and barley). The third part consists to identify the mitadin grains from hard wheat and to classify them in three categories of mitadinage.

Image matching based on LBP and SIFT descriptor

In this paper, we propose a new approach for extracting invariant feature from interest region. The new descriptor is inspired from the original descriptor SIFT (Scale Invariant Feature Transform) which is widely used in image matching by extracting interest points (IPs). However, this descriptor performs badly when the background is complex or corrupted with noise. Then, we adopt the local binary Pattern (LBP) descriptor with uniform pattern and the center symmetric local binary pattern (CSLBP) instead of a gradient feature used in the SIFT algorithm. To do so, we present new descriptors based on different combinations of SIFT, LBP and CSLBP descriptors to improve matching results. Thus, we compute different evaluation measures such as repeatability, recall and precision for various images transformations (blur attack, rotation and affine transformation). Experiments, which are achieved on two different databases, show that the descriptors leads to better results.

Non-Rigid Objects Recognition: Automatic Human Action Recognition in Video Sequences

Non-rigid objects recognition is an important problem in video analysis and understanding. It is nevertheless a challenging task to achieve due to the properties carried out by the nonrigid objects, and is more complicated by camera motion as well as background variation. Human body recognition in video sequences is the best application of the non-rigid objects recognition due to the large capacities of the human body in doing actions and poses. These difficulties prohibit practical attempts toward conceiving a robust global model for each action class. Human body recognition is highly interesting for a variety of applications: detecting relevant activities in surveillance video, summarizing and indexing video sequences. It relies, however, on the interpretation of the body movements and classifies them in different events.

Video Saliency Using Supervoxels

Physiology and neural systems researchers revealed that the visual system is attracted by some parts of an image more than others. Different computational models were developed to simulate the visual system. In this paper we propose a video saliency model that helps to predict and detect the regions of interest in each video frame. We use a supervoxel segmentation as an indicator of dynamic objects. Based on the observation that dynamic objects attract attention when an observer watches a video sequence, supervoxel segmentation provides a first estimation for what belongs to foreground and background. Then, a saliency score is attributed to each supervoxel according to its motion distinctiveness. Experiments over two benchmark datasets, using several evaluation metrics have shown that our proposed method outperforms five saliency detection methods.

Hybrid local and global descriptor enhanced with colour information

Feature extraction is one of the most important steps in computer vision tasks such as object recognition, image retrieval and image classification. It describes an image by a set of descriptors where the best one gives a high quality description and a low computation. In this study, the authors propose a novel descriptor called histogram of local and global features using speeded up robust feature (SURF) descriptor (HLGSURF) based on a combination of local features obtained by computation of Bag of words of SURF and global features issued from a novel operator called upper and lower local binary pattern (ULLBP) that encodes the texture analysis associated with wavelet transform. To enhance the effectiveness of the descriptor, the authors used the colour information. To evaluate the proposed method, the authors carried out experiments in different applications such as image retrieval and image classification. The performance of the suggested descriptor was evaluated by calculating both precision and recall values using the challenging Corel and COIL-100 datasets for image retrieval. For image classification, the performance was measured by the classification rate using the challenging Corel and MIT scene datasets. The experimental results showed that the proposed descriptor outperforms the existing state of the art results.

Salient regions detection method inspired from human visual system anatomy

Humans Region of Interest is considered as one of the most challenges problems in visual perception field. Due to the huge amount of information carried out, the research in the field of human region of interest tries to avoid the data overload by choosing salient areas from the total visual scene to be processed at first. In this context, salient regions detection becomes an important task to achieve. Inspired from the Humans Visual System anatomy, we introduce an algorithm to detect salient area in an image, achieved by merging four maps: color, intensity, orientation and central map. Our approach was evaluated using MSRA dataset and the result showed better performance compared to recent methods in this field.

Detecting and Recognizing Salient Object in Videos

Saliency detection has been an interesting research field. Some researchers consider it as a segmentation problem some others treat it differently. In this paper, we propose a novel video saliency framework that detects and recognizes the object of interest.

New matching method for human body tracking

In this paper we present a new method for interest points matching to realize human body tracking in video sequences. The developed algorithm combines direct and indirect similarity measures evaluated when applying luminosity variation and motion blur noises. This new approach considers different matching constraints such as: cross-matching, uniqueness constraint and interest points appearances and disappearances between consecutive images. The algorithm was evaluated on two different datasets and leads to high values of Good Tracking Rate.

Synthesis of spatio-temporal interest point detectors: Harris 3D, MoSIFT and SURF-MHI

The purpose of this paper is to evaluate and compare different spatio-temporal interest points (STIP) detectors which are considered as extensions of the most common interest points (IP) detectors in 2D images to 3D space-time features: Harris 3D, Motion Scale Invariant Feature Transform (MoSIFT) and Speeded Up Robust Features Motion History Image (SURF-MHI). This paper uses the criteria of repeatability and time execution to evaluate the approaches. We proposed to find matching using the approximation method of nearest neighbors. We investigate the performance of these methods for illumination variation, scale variation, rotation and compression were applied over three dataset. All the experiments were implemented in Matlab computing environment.