<?xml version="1.0" encoding="UTF-8"?><?xml-stylesheet type="text/xsl" href="static/style.xsl"?><OAI-PMH xmlns="http://www.openarchives.org/OAI/2.0/" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xsi:schemaLocation="http://www.openarchives.org/OAI/2.0/ http://www.openarchives.org/OAI/2.0/OAI-PMH.xsd"><responseDate>2026-11-21T07:03:39Z</responseDate><request verb="GetRecord" identifier="oai:macsphere.mcmaster.ca:11375/31624" metadataPrefix="dim">https://prod-ms-be.lib.mcmaster.ca/server/oai/request</request><GetRecord><record><header><identifier>oai:macsphere.mcmaster.ca:11375/31624</identifier><datestamp>2026-09-24T21:18:27Z</datestamp><setSpec>com_11375_271</setSpec><setSpec>col_11375_272</setSpec></header><metadata><dim:dim xmlns:dim="http://www.dspace.org/xmlns/dspace/dim" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:doc="http://www.lyncode.com/xoai" xsi:schemaLocation="http://www.dspace.org/xmlns/dspace/dim http://www.dspace.org/schema/dim.xsd">
   <dim:field mdschema="dc" element="contributor" qualifier="advisor">Ali, Emadi</dim:field>
   <dim:field mdschema="dc" element="contributor" qualifier="author">Saad, Sara</dim:field>
   <dim:field mdschema="dc" element="contributor" qualifier="department" lang="en_US">Electrical and Computer Engineering</dim:field>
   <dim:field mdschema="dc" element="date" qualifier="accessioned">2025-05-06T14:21:21Z</dim:field>
   <dim:field mdschema="dc" element="date" qualifier="available">2025-05-06T14:21:21Z</dim:field>
   <dim:field mdschema="dc" element="date" qualifier="issued">2025</dim:field>
   <dim:field mdschema="dc" element="identifier" qualifier="uri">http://hdl.handle.net/11375/31624</dim:field>
   <dim:field mdschema="dc" element="identifier" qualifier="uri">https://doi.org/10.71548/11868</dim:field>
   <dim:field mdschema="dc" element="description" qualifier="abstract" lang="en_US">Bias in automated face detection—where certain underrepresented groups consistently&#xd;
experience degraded or erroneous inference—remains a significant challenge in computer&#xd;
vision. This thesis proposes an efficient, fairness-aware face detection framework that&#xd;
mitigates bias across both demographic groups and other facial attributes, while leveraging&#xd;
GPU acceleration for real-time performance. The core contribution is a novel joint&#xd;
training approach that integrates a Variational Autoencoder (VAE) branch and a reinforcement&#xd;
learning–based rarity-driven resampling strategy into a pre-trained Faster&#xd;
R-CNN detector. This design enables the model to learn latent representations of underrepresented&#xd;
facial features (e.g., rare demographic or appearance traits) and rebalance&#xd;
the training process accordingly—notably without requiring explicit demographic or attribute&#xd;
labels. A rarity-aware sampling policy, guided by the VAE’s latent space and&#xd;
treated as a black-box optimization problem via reinforcement learning, preferentially&#xd;
up-samples infrequent examples during training to improve fairness for minority groups.&#xd;
Using face detection with a ResNet-50 backbone as a challenging testbed, we demonstrate&#xd;
that this approach consistently narrows the performance gap between well-represented&#xd;
and underrepresented groups without degrading overall accuracy. Trained on the CelebA&#xd;
dataset and evaluated on CelebA and FairFace benchmarks, the proposed model achieves&#xd;
significantly higher detection precision (mAP) and localization quality (mean IoU) compared&#xd;
to a baseline Faster R-CNN detector without VAE integration. Specifically, on&#xd;
CelebA, we observed improvement in mean IoU for all attributes (40 attribute groups)&#xd;
ranging from a minimum of +0.8% (Blurry, N = 1926, p &lt; 0.05) to a maximum of +5.3%&#xd;
(Rosy Cheeks, N = 3206, p &lt; 0.001). Similarly, on the FairFace dataset (14 race and gender subgroups), mean IoU improved by +1.3% (Female, N = 30001, p &lt; 0.001) up&#xd;
to +3.8% (Male, N = 31971, p &lt; 0.001). To ensure practical deployment, the trained&#xd;
model was exported via ONNX Runtime and optimized for GPU inference. On GPU, the&#xd;
ONNX-exported model attains nearly a 19× throughput speed-up over CPU execution,&#xd;
enabling real-time face detection. In summary, this research demonstrates a fast and&#xd;
fair face detection method that significantly improves performance on potentially biased&#xd;
datasets without sacrificing efficiency or accuracy.</dim:field>
   <dim:field mdschema="dc" element="description" qualifier="degreetype" lang="en_US">Thesis</dim:field>
   <dim:field mdschema="dc" element="description" qualifier="degree" lang="en_US">Master of Applied Science (MASc)</dim:field>
   <dim:field mdschema="dc" element="description" qualifier="layabstract" lang="en_US">Face detection systems often struggle to perform fairly across different people—detection&#xd;
can be less reliable for some races or genders. This thesis introduces a new face detector&#xd;
designed to work equitably. The system is trained on a large face dataset (CelebA) and&#xd;
evaluated on a balanced demographic dataset (FairFace), addressing biases in both demographics&#xd;
(race, gender) and other factors (like eyeglasses or lighting). The approach uses&#xd;
an advanced AI technique called a Variational Autoencoder, combined with reinforcement&#xd;
learning and a state-of-the-art face detection model, to recognize and compensate&#xd;
for underrepresented facial features. As a result, the system learns to avoid bias and improves&#xd;
detection accuracy for diverse groups. Importantly, it runs efficiently on modern&#xd;
graphics processors in real-time using ONNX Runtime, making the model deployable on&#xd;
different hardware without losing speed. In summary, this research demonstrates a fast&#xd;
and fair face detection method that better serves all demographic groups.</dim:field>
   <dim:field mdschema="dc" element="language" qualifier="iso" lang="en_US">en</dim:field>
   <dim:field mdschema="dc" element="title" lang="en_US">GPU ACCELERATED BIAS MITIGATION IN FACE DETECTION USING VARIATIONAL AUTO ENCODERS AND ONNX RUNTIME</dim:field>
   <dim:field mdschema="dc" element="type" lang="en_US">Thesis</dim:field>
   <dim:field mdschema="others" element="access-status">open.access</dim:field>
</dim:dim></metadata></record></GetRecord></OAI-PMH>