Dual-Region Preprocessing for Machine-Friendly JPEG Compression
Binzhe Li, Chae Eun Rhee · 2025
This paper proposes a region-of-interest (ROI)-based image preprocessing method to enhance JPEG compression for machine vision tasks. Unlike conventional approaches that apply preprocessing only to non–region-of-interest (NROI) areas, the proposed method additionally applies Gaussian blur to ROI regions to suppress noise and reduce compression artifacts. Experimental results on the MS COCO dataset with YOLOv5 demonstrate that the method achieves significant bitrate savings—up to 26.2%—while maintaining object detection accuracy. The approach is lightweight, fully compatible with standard JPEG codecs, and adaptable to real-time and edge computing environments.