Efficient smart distributed face identification using the MixMaxSim decision function

Abstract

Recognizing a large number of people is a common challenge in face identification applications, involving decreased accuracy, increased memory and time complexities. To address these issues, this study introduces a three-module approach: “toilers,” “affinity-meter,” and “decision-maker.” Unlike the random distribution methods used in previous solutions, this method employs clustering to distribute the problem into subnetworks called “toilers.” The toiler’s module calculates the likelihood of test data belonging to each class of each toiler, using the last layer outputs of deep learning models. Meanwhile, the affinity-meter module determines the similarity between the test data and the average of each class, employing a similarity measure. The decision-maker module combines the reports from the previous two modules and selects the final class, utilizing a mix of the max-max criterion and the similarity criterion. The proposed method outperforms existing solutions, achieving improved recall, precision, and F1-score. It effectively addresses memory, speed, and accuracy issues in face identification, surpassing both no-distribution and random methods on Glint360K, VGGFace2, and MS-Celeb-1M datasets. Overall, this method offers a more efficient and accurate approach by distributing the problem into subnetworks, demonstrating superior performance and scalability for large-scale face recognition applications.