Low-bitrate digital video often suffers from the artifact of texture flattening. Texture synthesis can be used to revive the
removed texture. Patch-based synthesis provides a quite general method for texture synthesis. However, this method still
requires a substantial bitrate to transmit example patches. We propose a method for stochastic texture synthesis which
requires only a very low bitrate (less than 1 kbit/sec) that can replace patch-based synthesis for random textures. Spatial
correlation is modeled as a 2-dimensional Moving Average (MA) process. To achieve a faithful representation of
temporal evolution, we use a translation+scaling motion model combined with a finite impulse response (FIR) filter.
Experiments show that we can successfully reduce texture flattening for a range of random textures such as grass and
In recent literature, privacy protection technologies for biometric templates were proposed. Among these is the so-called helper-data system (HDS) based on reliable component selection. In this paper we integrate this approach with face biometrics such that we achieve a system in which the templates are privacy protected, and multiple templates can be derived from the same facial image for the purpose of template renewability. Extracting binary feature vectors forms an essential step in this process. Using the FERET and Caltech databases, we show that this quantization step does not significantly degrade the classification performance compared to, for example, traditional correlation-based classifiers. The binary feature vectors are integrated in the HDS leading to a privacy protected facial recognition algorithm with acceptable FAR and FRR, provided that the intra-class variation is sufficiently small. This suggests that a controlled enrollment procedure with a sufficient number of enrollment measurements is required.
We propose a novel multistage facial feature extraction approach
using a combination of 'global' and 'local' techniques. At the
first stage, we use template matching, based on an
Edge-Orientation-Map for fast feature position estimation. Using
this result, a statistical framework applying the Active Shape
Model (ASM) is initialized and deformed to fit the real face
image. In our proposal, we use a 2-D pattern search-and-fitting
scheme guiding the deformation process, which provides more
robustness and faster convergence than the traditional ASM. Our
proposed approach for feature extraction shows good results
dealing with a test set composed of faces images which are quite
dissimilar with the faces used for the statistical training of the
face model. The convergence area of our proposed technique almost
quadruples compared to the ASM, while the amount of faces doubles
for which the convergence is reached. The total processing for
feature extraction takes less than 1 second for 250x250
face images on a Pentium-IV PC (1.7GHz).
In a home environment, video surveillance employing face detection and recognition is attractive for new applications. Facial feature (e.g. eyes and mouth) localization in the face is an essential task for face recognition because it constitutes an indispensable step for face geometry normalization. This paper presents a new and efficient feature localization approach for real-time personal surveillance applications with low-quality images. The proposed approach consists of three major steps: (1) self-adaptive iris tracing, which is preceded by a trace-point selection process with multiple initializations to overcome the local convergence problem, (2) eye structure verification using an eye template with limited deformation freedom, and (3) eye-pair selection based on a combination of metrics. We have tested our facial feature localization method on about 100 randomly selected face images from the AR database and 30 face images downloaded from the Internet. The results show that our approach achieves a correct detection rate of 96%. Since our eye-selection technique does not involve time-consuming deformation processes, it yields relatively fast processing. The proposed
algorithm has been successfully applied to a real-time home video surveillance system and proven to be an effective and computationally efficient face normalization method preceding the face recognition.
This paper concentrates on exploiting fast human face detection techniques for home video surveillance applications. The proposed method uses successive face detectors with incremental complexity and detection capability. The detectors are cascaded in such a way that each detector progressively restricts the possible face candidates into fewer areas. The proposed detectors, listed in the order of usage and complexity, are: (1) skin-color detector, (2) face structure detector which uses probability-based facial feature verification, and (3) three parallel learning-based detectors which take several representations of face candidates as inputs. The
adopted representations are the pixel representation, the partial profile representation and the eigenface representation. The initial pruning of large areas of non-face regions significantly decreases the number of input windows for the learning-based face detector. This largely reduces the high computation cost for most learning-based detection approaches, while retaining the high detection accuracy and learning capabilities. Experimental results show that our proposal achieves an average of 0.3 - 0.4 second per frame processing speed with an image resolution of 320 by 240 pixels. An average of 92% detection rate is achieved for a test set composed of downloaded photos, standard test sequences and self-made sequences.