Assessment and Estimation of Face Detection Performance Based on Deep Learning for Forensic Applications

Area of Science:

• Computer Vision
• Forensic Science
• Artificial Intelligence

Background:

• Face recognition is crucial for criminal investigations, aiding in identifying fugitives and in cases of child sexual abuse.
• Deep learning face detectors offer high accuracy but demand substantial computational resources and processing time.
• Real-world forensic applications necessitate face recognition systems that handle low-quality images and meet real-time processing demands.

Purpose of the Study:

• To evaluate the speed-accuracy trade-off of popular deep learning face detectors on diverse datasets and hardware.
• To develop a predictive model for estimating the performance of face detection systems.
• To provide a practical tool for forensic laboratories to optimize face detection strategies.

Main Methods:

• Assessed three deep learning face detectors on WIDER Face and UFDD datasets across various CPUs and GPUs.
• Investigated the impact of image resizing on detection speed and accuracy.
• Developed a multiple linear regression model to predict performance metrics (speed and accuracy).

Main Results:

• The optimal speed-accuracy trade-off was achieved by resizing images to 50% on GPUs and 25% on CPUs.
• The developed regression model demonstrated a Mean Absolute Error (MAE) of 0.113 in performance estimation.
• This predictive capability is highly promising for practical forensic applications.

Conclusions:

• Image resizing is a critical factor in optimizing deep learning face detection for forensic use.
• A reliable regression model can accurately estimate face detection performance, aiding forensic practitioners.
• The findings offer valuable insights for deploying efficient and effective face recognition tools in forensic laboratories.