Forensic License Plate Recognition with Compression-Informed Transformers

TL;DR

This work addresses forensic license plate recognition under severe compression by introducing a compression-informed Transformer that uses a side input $c_n$ representing JPEG quality factors. The model processes column-wise image slices and outputs a character sequence, embedding the compression cue into the Transformer to guide recognition under degradation. Empirical results show strong performance on real-world data with a compact model and substantial gains on synthetic, highly degraded data when using a $QF$-aware knowledge embedding (optimal around $K=50$ classes). The approach yields competitive accuracy with far fewer parameters than prior methods and demonstrates the importance of incorporating compression metadata for robust FLPR in surveillance scenarios. The SynthGLP dataset enables thorough analysis of degradation effects and highlights practical implications for forensic image analysis and law-enforcement workflows.

Abstract

Forensic license plate recognition (FLPR) remains an open challenge in legal contexts such as criminal investigations, where unreadable license plates (LPs) need to be deciphered from highly compressed and/or low resolution footage, e.g., from surveillance cameras. In this work, we propose a side-informed Transformer architecture that embeds knowledge on the input compression level to improve recognition under strong compression. We show the effectiveness of Transformers for license plate recognition (LPR) on a low-quality real-world dataset. We also provide a synthetic dataset that includes strongly degraded, illegible LP images and analyze the impact of knowledge embedding on it. The network outperforms existing FLPR methods and standard state-of-the art image recognition models while requiring less parameters. For the severest degraded images, we can improve recognition by up to 8.9 percent points.

Figures (4)

• Figure 1: The proposed Transformer architecture with knowledge embedding.
• Figure 2: Samples from featured real world and synthetic sets
• Figure 3: acclp results averaged over $c_n \in \mathcal{C}_1$ on SynthGLPTFull. models perform comparably well for $r_w \ge 30$ and surpass all CNN methods. The advantage of knowledge embedding steadily enlarges with decreasing $r_w$ and is highest for $r_w = 20$ as depicted in Tab. \ref{['tab:res_20']}.
• Figure 4: acclp for the best performing baseline CRNN moussa2021sequence and our method.