Performance Trade-Off Analysis of Faster R-CNN with Grid-Based Histogram for Student Face Detection
TL;DR
The novelty of this study lies in applying a Grid-Based Histogram as a lightweight replacement for conventional anchor generation in Faster R-CNN, resulting in a preliminary speed–accuracy trade-off rather than an overall performance improvement.
Abstract
Face detection supports applications such as security, identity management, human–computer interaction, and academic information systems. Faster R-CNN is known for strong detection accuracy, but its Region Proposal Network produces many candidate boxes, including background proposals, which may increase processing cost. This study replaces the conventional anchor-generation process with a Grid-Based Histogram method to improve inference efficiency while retaining competitive detection performance. Experiments were conducted on 1132 annotated student profile images collected from the Campus Information System of Institut Teknologi Del. The standard and modified models were evaluated using mean Intersection over Union (IoU), mean Average Precision at IoU 0.50 (mAP@50), and average latency per image with an inference batch size of one. Standard Faster R-CNN achieved an IoU of 0.7595, an mAP@50 of 0.9818, and a latency of 0.170 s per image. The modified model obtained an IoU of 0.7519, an mAP@50 of 0.9719, and a latency of 0.147 s per image. Thus, latency decreased by about 13.53%, with small reductions in localization and detection accuracy. The novelty of this study lies in applying a Grid-Based Histogram as a lightweight replacement for conventional anchor generation in Faster R-CNN, resulting in a preliminary speed–accuracy trade-off rather than an overall performance improvement.