IntegraScale: A Unified Integral Framework for PSF-Aware Quantized Subpixel Image Scaling

Jeongseok Kim, Daeyeon Kim, Eunjin Koh, Daeyoung Kim

Abstract: To develop detection and tracking of small, distant objects, data augmentation requires an accurate image-synthesis method that can downsample high-resolution targets to arbitrary scale factors and place them with subpixel precision. Existing approaches decompose downscaling, translation, and blurring into multiple stages, which 1) induce up to 0.5-pixel localization error at subpixel positions, 2) cause energy loss due to information loss at each stage, and 3) increase computational cost while making high-order filters sensitive to quantization. We propose IntegraScale, a unified interpolation framework that simultaneously performs scaling, translation, and point spread function (PSF) application within a single integral formulation. By performing integral computations with an analytic overlap function ( $f_{O}$ ) and a Gaussian function ( $f_{G}$ ), the method faithfully reproduces ideal low-pass behavior and accurate sampling. On comprehensive experiments with $200\times 200$ input images, the proposed method shows strong performance across key metrics. In terms of positional accuracy, the subpixel error of IntegraScale is nearly zero, effectively removing the systematic bias observed in prior methods; at a reduction ratio of $0.057\times $ , it achieves $28.5\times $ lower error than the conventional method (Conv). In efficiency, it is slightly slower than Conv but about $3.5\times $ faster than Lanczos at comparable accuracy. Under input-only quantization at 16/8/4-bit and full 8-bit quantization of the entire pipeline, the centroid MAE remains $\leq 0.25$ pixels across all tested settings. IntegraScale executes a single, PSF-aware discrete integral that eliminates intermediate re-quantization and preserves energy. We also validate an INT8 (W8A8/INT32) path maintaining subpixel accuracy. Overall, through single-pass integral processing and an analytic computational structure, IntegraScale prevents error accumulation and ensures numerical stability, making it readily applicable as a precise image synthesis technique even in resource-constrained settings such as edge computing.

External IDs:dblp:journals/access/KimKKK25