Performance Analysis of Haar Cascade-Based Face Detection in Multi-Face Images under Diverse Compression Algorithms

With the tremendous development of face detection systems, there is a growing need to achieve highly accurate detection results in images compressed using different compression algorithms.This manuscript provides face detection analysis in images representing different numbers of faces (1, 3, 5, and 10 faces) from both frontal and non-frontal perspectives.The images extracted from the FDD (Face Detection Dataset) database were compressed using three different compression algorithms -JPEG, JPEG2000, and SPIHT, for different bits-per-pixel values.The analysis was performed by using the Haar Cascade Classifier, implemented in Python.The quality of face detections was determined using the objective measures: F-measure (based on reference values from the GroundTruth images) and Det.F (number of detected faces).Based on the results presented in the tables, it can be concluded that face detection behaves slightly differently depending on the value of bits-per-pixel and the applied compression algorithm, but vastly differently depending on the angle of perspective.The Haar Cascade Classifier has proven to be the best solution when it is necessary to perform face detection in compressed, frontal face images, especially for a small number of faces; non-frontal images with a large number of faces have proven to be the most challenging assignment for the Haar Cascade Classifier.