Effects of Conventional, Adaptive, and AI-Powered Upscaling on the Legibility of License Plates
arXiv:2611.18949
Abstract
Small details in digital images -- a face, a street sign, a license plate -- are routinely upscaled in hopes of making them interpretable. Upscaling options now range from conventional interpolation, whose operations are spatially invariant and fully explainable, to AI-powered methods that synthesize detail potentially absent from the source. Here, we ask which of these techniques most reliably helps human readers. In two online experiments, we measured legibility of 255 US license plates ranging in resolution from 128 to 8 pixels tall, as well as their legibility after upscaling by eight different methods. Legibility of the un-upscaled plates collapsed from ceiling to floor across a narrow band of plate heights (32 to 10 pixels). Conventional upscaling techniques (bilinear, bicubic, and Lanczos interpolation) gave modest gains of about 5-9 percentage points within this range. Nearest-neighbor interpolation, however, was harmful, rendering unreadable many plates that were readable without any upscaling. Advanced methods diverged: SwinIR, an open-source AI enhancer, raised performance by up to 38 points relative to no upscaling, while an adaptive denoiser-based image prior and commercial AI upscaler tended to impair legibility. We found no strong evidence of consistent character hallucinations across methods. Decomposing the task, conventional interpolators mainly helped readers count characters, whereas SwinIR alone made the characters themselves more identifiable.