A. Vinay

Expediting Automated Face Recognition Using the Novel ORB 2 -IPR Framework

Face Recognition (FR) is at the forefront of distinctly unresolved challenges in the domain of Computer Vision, due to the sharp accuracy and performance drops it undergoes, when there are pronounced variations in parameters such as illumination, pose, background clutter and so on between the input and database faces. In this paper, we attempt to expedite the performance of automated FR with real-time images, using a novel framework called ORB2-IPR (ORB based Bag of Interest Points using RANSAC), which exhaustively learns a vocabulary of highly discriminative facial interest points from the facial database images (which can be referred to, and compared directly, instead of following the conventional time-intensive approach of comparing a given input face with each database face separately) by employing the cost-effective ORB (Oriented Fast Rotated Brief) descriptor (instead of the commonly employed SIFT and SURF descriptors), followed by the application of RANSAC (Random Sample Consensus) as a post-processing step to remove noise in the form of outliers, in order to improve the accuracy of the system. We will conclusively demonstrate that our technique is capable of rendering superior performance than the state-of-the-art methodologies using extensive mathematical arguments and by carrying out ample experimentations on the benchmark ORL, Face 95 and LFW databases.

Face Recognition Using the Novel Fuzzy-GIST Mechanism

Face Recognition (FR) is one of the most thriving fields of contemporary research, and despite its universal application in authentication and verification systems, ensuring its effectiveness in unconstrained scenarios has predominantly remained an on-going challenge in Computer Vision, because FR systems experience considerable loss in performance, when there exists significant variation between the test and database faces in terms of attributes such as Pose, Camera Angle, Illumination and so on. The potency of FR systems markedly declines in the presence of noise in a given face and furthermore, the performance is also determined to a large degree by the Feature Extraction technique that is employed. Hence in this paper, we propose a novel mechanism known as Fuzzy-GIST, that can proficiently perform FR by adeptly handling real-time images (which contain the aforementioned unconstrained attributes) in low-powered portable devices by employing Fuzzy Filters to eliminate extraneous noise in the facial image, prior to feature extraction using the computationally less demanding GIST descriptor. Backed by relevant mathematical defense, we will establish the efficacy of our proposed system by conducting detailed experimentations on the ORL and IIT-K databases.

Dominant feature based convolutional neural network for faces in videos

Standard face recognition modules are fabricated for general-purpose applications while few have been designed with speed in mind. This paper proposes an efficient architecture for face recognition in which two self-contained Convolutional Neural Networks (CNNs) are used to detect and recognize faces in regions containing a dense grouping of Features from Accelerated Segment Test (FAST). This configuration proves to be practical for videos as it is selective in its analysis of an input frame. City surveillance and public safety is a critical issue in smart cities and the deployment of Smart Video Surveillance systems is the need of the hour. Typically, the problem at hand will be person identification which is the association of a biometric trait with a particular human being. FAST key points can be generated and analyzed in near real-time and that data can be used to extract and process faces in the background. The CNNs were trained using a combination of datasets of labelled faces, videos and trivial objects. The results obtained upon analyzing the performance of the system on the ChokePoint dataset proved very insightful. This configuration leads to a very effective face recognition system.

RISA: Rotation Illumination Scale and Affine Invariant Face Recognition

Face Recognition (FR) has been on the forefront of research efforts for the past two decades. In spite of considerable strides, it still suffers from the curse of false matches in the presence of variations in terms of parameters such as affine, scale, rotation and illumination. Since, real world images inherently consists of such variations, an effective FR system, should handle such variations deftly. Hence, in this paper, we propose a robust, yet simple and cost effective technique for overcoming some of the aforementioned challenges. The first stage of the proposed system deals with illumination variations by performing logarithm transform on the input face images. Further, the Non-subsampled Contourlet Transform (NSCT) is used to decompose the logarithm transformed facial images into low frequency and high frequency components. Subsequently, histogram equalization is carried out on the low frequency components. Finally, we employ Affine Scale Invariant Feature Transform (ASIFT) to find corresponding points that are translation and scale invariant. We will demonstrate by carrying out extensive experimentations on the benchmark datasets: ORL, Grimace, Face95 and Yale, that the proposed technique is more robust and yields comparable efficacy to most of the contemporary approaches.

A Sparse and Inlier Based SIFT and SURF Features for Automated Face Recognition

Face Recognition (FR) is a prolific form of biometric that has spawned a myriad of inventive applications for commercial and law-enforcement scenarios and has fostered several novel research directions. In the FR process, the choice of the feature extractor governs the overall efficiency and in that regard, SIFT and SURF are two prominent feature extraction mechanisms that are frequently employed due to their robustness with respect to scale, translation, illumination and rotation. Even though the SIFT and SURF descriptors are immensely effective, they are cluttered with redundant key-points and noise, which we aim to tackle by employing the Sparse Singular Value Decomposition (SSVD) method to perform Dimensionality Reduction, and RANSAC to remove noise in the form of outliers. In this paper, we will conclusively demonstrate by utilizing extensive mathematical arguments, and by performing exhaustive experimentations over the benchmark ORL and LFPW databases, that the proposed SIFT-SSVD-RANSAC and SURF-SSVD-RANSAC methodologies are more effective than their classical counterparts, as they are capable of handling extreme variations between the matched images in terms of scale, zoom, view-point and so on. The findings proffered by this work, coupled with our other studies, form a series intended to aid developers in making prudent decisions in order to build proficient FR systems.

Face Recognition using VLAD and its Variants

Face Recognition (FR) has grown to be one of the most productive fields in the domain of Computer Vision (CV) due to the wide range of applications it has bestowed in commercial and law enforcement settings and remains one of the most formidable problems in CV. One of the central issues with FR is that its recognition performance tends to suffer when the size of the database to search through is large, due to the cost of needing to generate and compare descriptors for each individual face. Hence in this paper, we address this issue by employing VLAD (Vector of Locally Aggregated Descriptors) to aggregate local face descriptors (which are pooled using Bag-of-Words (BOW) algorithm) and classified using RBF-SVM (Radial Basis Function-Support Vector Machine) to output a predicted label file, which can be used to directly access the descriptors, instead of computing them from individual face images for each comparison. Although variants of VLAD exist, they have been proposed primarily for object recognition and hence we proffer three novel VLAD variants for FR by utilizing the popular feature descriptors: SIFT (Scale Invariant Feature Transform), SURF (Speeded Up Robust Features) and ORB (Oriented Fast Rotated Brief). We will demonstrate using comprehensive experimentations on the ORL database that the proposed variants can considerably improve the performance of FR over the contemporary state-of-the-art algorithms.