Chinese Researchers Detail Improved ‘Particle Swarm Optimization’ Technique for Facial Recognition

Researchers at Guangdong Songshan Polytechnic have detailed a new technique that they say can help facial recognition systems attain a better balance between speed and accuracy. The technique involves the use of a “particle swarm optimization algorithm”: Essentially, it iteratively explores different ways of configuring face images to get the clearest picture until it finds the best option in terms of accuracy and efficiency.

The implementation process of the Facial Compensation with Improved PSO (FCAI) algorithm includes multiple steps. First, the face dataset is converted to grayscale images and imported. Next, the system needs to figure out some special values or “factors” that help it recognize faces better. These factors are called S1, S2, S3, up to S8. Think of them as unique characteristics or measurements of each face, like the distance between the eyes or the shape of the nose. The system uses specific formulas or mathematical rules to calculate these factors for each face in the dataset.

The improved PSO algorithm analyzes all the factors S1-S8 and their values for each face. Then, it figures out the best “compensation coefficients” to adjust these factors, just like fine-tuning the settings of a camera to take a clearer picture. These coefficients are like secret codes that help improve the face recognition system’s accuracy.

Once the system has the best compensation coefficients, it creates a special image called a “feature descriptor” for each original face picture. This image summarizes all the important facial features, like a unique face signature, based on the factors and coefficients found in the previous steps.

The system generates a feature descriptor image for each original image, and Principal Component Analysis (PCA) is used to reduce the dimensionality of the image’s histogram. This is a way to reduce the amount of information in the feature descriptor image, while keeping the most important details.

Finally, a Support Vector Machine (SVM) algorithm learns from the simplified face summaries (the PCA results) and creates a special model that can recognize faces.

To optimize the combination of compensation coefficients, the improved PSO algorithm takes inspiration from bird flock behavior, updating particle positions and velocities iteratively. It considers the historical best position of the particles to avoid getting trapped in local optima and dynamically adjusts its parameters during the iteration process. In other words, it keeps exploring different possibilities until it finds the best combination of factors that give both high accuracy and speed. The system can adapt to different situations and problems, just like how a car adjusts its speed and direction to handle different roads.

The technology was experimentally validated on three datasets: ORL, YALE, and MU_PIE. The results demonstrated the effectiveness of the FCAI algorithm, outperforming other widely used face recognition algorithms in terms of accuracy and generalizability, the researchers say.

Their full paper, including graphs and formulae, can be found on Nature.

Source: Nature

–

July 31, 2023 – by the FindBiometrics Editorial Team