Research

Facial Biometrics

• Shortly, Biometrics refers to the analysis and extraction of specific and distinguished biological and behavioural characteristics of an individual aiming at automatic recognition of individual. When the biological feature concerns face analysis, we call this facial biometrics. Unlike many biometrics data, the face recognition issue has its great advantage of being nonintrusive and non-interactive. More precisely, the task of face recognition does not require any physical contact, and does not necessarily involve end user interaction. Two major open issues still  exist:

  1. Face Presentation Attack Detection (F – PAD). An impostor can easily bypass a facial biometric based system by simply presenting a copy (still image, video, mask, etc.) of biometric data of a legitimate user in front of the camera sensor. The attempt of breaking the biometric system using such method is named spoofing attack. Typically, the reported performance of the state-of-the art (SoTA) approaches are assessed using publicly available data sets. Modern approaches include advanced AI techniques, including machine learning (ML) and deep learning (DL) methods.  Recent studies indicate poor to moderate performance of SoTA trained on specific data set and tested on a different set (cross database tests), where distribution varies (different attack devices with different parameters, variability in acquiring conditions, etc.). Moreover, there is a large gap between laboratory experiments and real-life condition performances. Therefore, F – PAD is still an active research area.

  2. Weak (cryptographic) security. Typically, a facial biometric system does not store the facial images. Instead, it applies complex mathematical processes to derive a facial signature, called biometric templates. When the biometric templates are stored as plaintext, an impostor may directly use the templates to access a non authorised account that could be permanently compromised (security issue). Moreover, this may lead to privacy leak (privacy issue). One direct approach would be to encrypt the templates using hash functions, similar to the traditional encryption approach used for passcodes. Such approach requires the input data (passcode) remain exactly the same in order to work. However, there are never two exact biometric templates due to many variables (face occlusion, different illumination acquiring conditions, face pose, noise, etc). A small change in pixel value leads to a large change in hash value (avalanche effect), causing unacceptable drop in the biometric system’s performance.

Image and Video Encoding and Representation

• While the Human Visual System efficiently compresses, encodes and represents the visual information, emulating its functionality by developing a biologically inspired artificial visual system is a hard task. Particularly, the analysis of visual information has attracted a specific interest as processing such very large amount of data is challenging. Neurophysiological investigations inspired the design of different models for image understanding and representation, mainly falling into one of the two categories: holistic or sparse. Modern approaches admit a hierarchical representation. However, many mechanisms have yet to be further developed and incorporated to mimic the HVS. Saliency – guided sparse image representation and compression is an excited topic along with temporal encoding of video data. The sparse representation concept looks very appealing as many surrounding video data contain redundant information which is discarded for further processing steps. Nonlinear decomposition of data (kernel based, for example) into sparse and independent components resembles some properties the biological receptive field and the firing rate. However, some redundant and overlapping correlated information might be necessary to have a robust model of the visual system for handling partially occluded object recognition tasks, for instance. Many contradictory constraints make the design of this system to be difficult.

Medical Data Analysis

• Automatic tumour detection from mamogramms may help radiologists in having a correct decision. Computer-aided detection systems cannot replace the medical personnel but they could highlight the abnormalities to help the radiologist detect early breast cancer, by speeding up the detection process. One of the most difficult tumour types to be detected, either automatic or by human (radiologist) visual inspection is the architectural distortion (AD). The lack of any regular pattern class for this tumour type makes the application of standard pattern recognition techniques to be almost useless. For AD, the best current (commercial) CAD systems are able to detect the tumour at a rate of less than 30 % accuracy, which is unacceptable.

• Automatic abnormal region of interest (ROI) detection, measurement and classification of lung tissue affected by COVID – 19 from CT images or X – Ray.