CNN-based thermal IR facial recognition systems for biometric identification

Many facial recognition systems rely on images captured in visible light, which can be affected by various factors such as lighting conditions, weather, and time of day, leading to fluctuations in identification rates. Moreover, differences in skin tones among ethnic groups can also impact the way light is reflected from their faces.

To address these challenges, thermal infrared facial recognition systems have been developed. These systems use IR cameras to capture the thermal emissivity of the facial surface, resulting in more stable images. Because thermal IR sensors detect the heat patterns emitted by human faces, they are not influenced by ambient lighting conditions.

While biometric identity applications and law enforcement agencies are exploring the potential use of thermal IR facial recognition systems, it’s crucial to acknowledge the obstacles involved. Thermal images may be susceptible to issues such as noise, blurring, loss of spatial resolution, and temperature variations.

A team of researchers associated with Arab Open University and Kuwait Technical College has suggested using convolutional neural networks (CNN) to recognize thermal face images with degradation. The outcomes demonstrate that the CNN model exhibits strong recognition capability and performs effectively with such images.

Mechanics of CNN

Convolutional neural networks leverage supervised learning, where they are trained on labeled datasets with known input-output pairs. Throughout the training process, CNNs learn the most effective filters and weights for each layer by minimizing the disparity between predicted and actual labels using backpropagation and gradient descent.

The research paper highlights the use of stochastic gradient descent, an iterative optimization algorithm, to minimize a function by adjusting the model parameters in a way that reduces the loss. By iteratively applying stochastic gradient descent and updating the parameters, CNNs refine their predictions. Typically, the final prediction for a given input is the class with the highest probability.

The study utilizes the ResNet-50 architecture, which is constructed by stacking residual blocks. Each block contains layers that execute 3×3 convolution, followed by batch normalization and activation functions. ResNet also addresses the challenge of the vanishing gradient, where the gradient diminishes significantly in the deeper layers of a network, hindering effective learning.

The experiments are carried out using a dataset of 7,500 thermal images. The scenarios encompass varying image quality levels, reduced spatial resolution, and differences in facial pose and expression. As the research indicates, the proposed CNN approach demonstrates high identification rates.

Article Topics

biometric identification  |  biometrics  |  facial recognition  |  infrared  |  thermal imaging