FEASIBILITY OF FACE RECOGNITION ALGORITHMS ON EDGE

Abstract

The emergence of IoT and its rapid growth increased significance of edge computing. Edge computing is efficient and necessary when the network bandwidth is low and volume of data that needs to be transmitted is large and could be used when IoT involves data images. It would be more efficient to bring computer vision to edge node rather than sending the image to the server. The main objective of this paper was to analyze main face recognition algorithms and compare their feasibility and accuracy on edge. They are: Eigenfaces, Fisherfaces and Local Binary Patterns algorithm. The algorithms were implemented and executed on Raspberry Pi as it is a device that is used often in IoT and edge computing. The algorithms were implemented using OpenCV library and Python programming language and tested on a custom dataset of 150 images of 10 people. Algorithms’ peformance was analyzed in terms of model training time, recognition time and recognition accuracy. The results showed the feasibility of the algorithms on edge. Fisherfaces algorithm recognized subjects 16 times faster than Eigenfaces and 5 times faster than LBPH algorithm, while providing slightly better recognition accuracy than both. Thus, it was concluded that Fisherfaces face recognition algorithm is the most suitable algorithm in terms of feasibility on Raspberry Pi edge device.