Hala M. Ebeid

Fusing directional wavelet local binary pattern and moments for human action recognition

Recently, transformation-based methods have been widely used in many computer vision areas because of their powerful representation ability. One of the most widely used transforms is the wavelet transform that has proved to be very useful in many applications. In this study, a new method for human action representation and description is proposed. This method combines the advantages of local and global descriptions. The method works by fusing the Hu invariant moments as global descriptors with a new local descriptor that is based on three-dimensional stationary wavelet transform and the concept of local binary patterns. The performance of the new method was examined in two different ways. The first one is by fusing the proposed directional global and local features in one feature vector, while the other is using the features of different directional bands separately to train multiple classifiers and then using a voting scheme to vote for the best match. The performance of the proposed method is verified using standard datasets, achieving high accuracy in comparison with state-of-the-art methods. In addition, the proposed method is proved to be robust to the changes in lighting and scale variations, but it exhibits limitations towards dynamic backgrounds.

Using MLP and RBF neural networks for face recognition: An insightful comparative case study

In this paper, two architectures neural network (NN) classifier models have been compared, multilayer perceptron (MLP) neural network with back-propagation algorithm and radial basis function (RBF) neural network. Capabilities of the presented architectures have been compared. The feature projection vectors, obtained through the Principal Component Analysis or called Eigenfaces method, are used as the input vectors for the training and testing of both NN architectures. Several factors affect the recognition performance; experimental results are applied to the ORL database which contains variability in expression, pose, and facial details. The experimental result showed that the Eigenfaces/RBF system has recognition error rates that are lower than those of the Eigenfaces/MLP system by 3%. Thus the Eigenfaces/RBF system performs better than the Eigenfaces/MLP system in terms of correct recognition rates and training convergence speed of the network.

Action Recognition Using Stationary Wavelet-Based Motion Images

Human action recognition is one of the most important fields in computer vision, because of the large number of applications that employ action recognition. Many techniques have been proposed for representing and classifying actions; yet these tasks are still non-trivial due to a number of challenges and characteristics. In this paper, a new action representation method is proposed. The proposed method utilizes the 3D Stationary Wavelet Analysis to encode the spatio-temporal characteristics of the motion available in the video sequences in a way similar to motion history images. The proposed representation was tested using Weizmann dataset, exhibiting promising results when compared to the existing state – of – the – art methods.

Directional Stationary Wavelet-Based Representation for Human Action Classification

This paper proposes a directional wavelet-based representation of natural human actions in realistic videos. This task is very important for human action recognition, which has become one of the most important fields in computer vision. Its importance comes from the large number of applications that employ human action classification and recognition. The proposed method utilizes the 3D Stationary Wavelet Analysis to encode the directional spatio-temporal characteristics of the motion available in video sequences. It was tested using the Weizmann dataset, and produced promising preliminary results (92.47 % classification accuracy) when compared to existing state–of–the–art methods.

Clustering-based Image Segmentation using Automatic GrabCut

GrabCut is one of the most powerful semi-automatic segmentation techniques. One main drawback of GrabCut is the need for user interaction in order to initialize the segmentation process. User interaction involves dragging a rectangle around the object of interest in the image to extract it. This restricts GrabCut for binary-label segmentation, where the image cannot be segmented into more than two; foreground and background segments. Unsupervised clustering is a powerful automatic tool for dividing images into a specified number of regions based on defined image features. The authors had introduced the SOFM clustering technique for GrabCut automation as a replacement to the user interaction. In this paper, K-means and Fuzzy C-means are introduced as new clustering techniques to automate GrabCut and improve the segmentation accuracy. Experimental results and comparisons with the previous results are carried out to test the efficiency of the proposed techniques in terms of segmentation quality and accuracy.

A method for contactless palm ROI extraction

Palmprint can be extracted from a hand using a low-cost webcam in a contactless manner. Using a webcam makes the enrollment process fast and convenient for users. Being contactless solve the hygiene issue and avoid copying the latent prints from sensors surface. However, a number of challenges arise in such environment; geometric transformations, the existence of finger rings, hand accessories, and other false objects. This paper proposes a palm ROI extraction method that is robust to these challenges. The method is based on blob analysis, morphological and geometrical operations without a need to pre-train or parameter adjustment. Its tested on three available hand databases that cover these challenges; namely, Sfax, IITD and PolyU 3D/2D. The palm ROI is considered to be wrongly extracted if it contains part of the background. The method achieves an extraction error of 0%, 0.27% and 0.26% for the three DBs, respectively. Applying a massive rotation and scaling tests leads to a minor increase in the extraction errors by 0.24%, 0.35% and 0.84% for the three DBs, respectively.

3D Mesh Segmentation Based on Unsupervised Clustering

3D mesh segmentation is considered an important process in the field of computer graphics. It is a fundamental process in different applications such as shape reconstruction in reverse engineering, 3D models retrieval, and CAD/CAM applications, etc. It consists of subdividing a polygonal surface into patches of uniform properties either from a geometrical point of view or from a perceptual/semantic point of view. In this paper, unsupervised clustering techniques for the 3D mesh segmentation problem are introduced. The K-means and the Fuzzy C-means (FCM) clustering techniques are selected for the development of the proposed clustering-based 3D mesh segmentation techniques. Since the mesh faces are considered the main element, the clustering technique is applied to the dual mesh. The 3D Euclidean distance is used as the distance measure to compute matching between mesh elements. Based on empirical results on a benchmark dataset of 3D mesh models, the FCM-based mesh segmentation technique outperforms the K-means-based one in terms of accuracy and consistency with human segmentations.

An unsupervised method for face photo-sketch synthesis and recognition

Face recognition is considered one of the most essential applications of Biometrics for personal identification. Face sketch recognition is a special case of face recognition, and it is very important for forensic applications. In this paper, we propose an unsupervised method for face photo-sketch recognition by synthesizing a pseudo-sketch from a single photo. The proposed method is the first unsupervised method that deals with face sketch recognition. The proposed photo-sketch synthesis step consists of two main steps, namely: edge detection and hair detection, which are applied on the grayscale image of the photo image. In the recognition step, the artist sketch is compared with the generated pseudo-sketch. PCA and LDA are used to extract features from the sketch images. The k-nearest neighbor classifier with Euclidean distance is used in the classification step. We use the CUHK database to test the performance of the proposed Method. Results for the synthesized sketches are compared with state-of-the-art methods, e.g., Local Linear Embedding (LLE) and Eigen transformation. The experimental results show that the proposed method generates a clear synthesis sketch and it defines persons more accurate than other methods. Moreover, in the recognition step, the proposed method achieves a recognition rate at the 1-nearest neighbor (rank1: first-match) range from 82% with PCA to 94% with LDA. The highest recognition rate is obtained at the 5-nearest neighbor (rank 5) is 98% that is better than some of the state-of-the-art methods.

Adaptive Multiple Kernel Self-organizing Maps for Hyperspectral Image Classification

Classification of hyperspectral images is a hot topic in remote sensing field because of its immense dimensionality. Several machine learning approaches had been effectively proposed for hyperspectral image processing. Multiple kernel learning (MKL) approaches are the most used techniques that have been promoted to improve the adaptability of kernel based learning machine. In this paper, an adaptive MKL approach is promoted for the classification of hyperspectral imagery problem. The core idea in the introduced algorithm is to optimize the convex combinations of the given base kernels during the training process of Self-Organizing Maps. Diverse types of kernel functions are used. The performance of the classifier based on the choice of the kernel function and its variables, Benchmark hyperspectral datasets are used. The experimental results demonstrate that the introduced MKLSOM learning algorithm gives a comparative solution to the state-of-the-art algorithms.

A Mean Features Method for Face Photo-Sketch Synthesis and Recognition

Converting a photo image to sketch, or conversely, is an essential step in face-sketch recognition. In this paper, we propose an efficient mean feature method to synthesize a sketch from a photo and vice versa. The main idea is to map a photo to the same sketch texture and vice versa. This is done by generating a mean features image from the training set. We used pseudo-sketch to sketch recognition as a performance measure for the proposed method. SIFT feature and Euclidean distance are used in the recognition step. We used CUHK viewed-sketch and PRIP-HDC forensic sketch databases in our experiments. Also, comparisons with state-of-the-art methods are presented. Experimental results for the CUHK database showed that the proposed method outperform some state-of-the-art method. We obtained a recognition rate of 96% at rank 1, which is better than some of the state-of-the-art methods. Our results for the PRIP-HDC database show improvement in the recognition rate from 34% to 57% at rank 50