| Title |
TFD-LP: Template-free Occlusion-aware License Plate De-identification via Adaptive Feature Fusion |
| Authors |
안민찬(Minchan Ahn) ; 이성우(Xingyou Li) ; 김학일(Hakil Kim) ; 권장우(Jangwoo Kwon) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.6.50 |
| Keywords |
De-identification; Generative adversarial network; Occlusion; Face swap adaptation; License plate |
| Abstract |
License plate de-identification aims to replace the original text while preserving visual attributes such as background and style. Existing GAN-based methods rely on standard plate templates and have limitations under occlusion. We propose TFD-LP, which adapts the Adaptive Feature Fusion Attention (AFFA) mechanism of the face swap architecture FaceDancer to the license plate domain, enabling template-free de-identification with specified target text. During this domain adaptation, we discovered that the generator produces adversarial noise?visually imperceptible patterns that fool the frozen OCR judge but remain unreadable to humans?caused by spatial information loss in the frozen OCR and AdaIN, and by the unconditional discriminator's inability to verify text content. To address these issues, we introduce a projection-based conditional discriminator for text-image conformity verification, and spatially-adaptive modulation with spatial-preserving encoding for spatial information enhancement. We further extend the framework to occluded plates via MASK tokens, masked OCR loss, and pattern-wise group sampling. On the CCPD dataset, TFD-LP achieves a De-ID Rate of 100%, LPIPS of 0.187, FID of 76.0, and CRNN 1-NED of 0.954. Ablation studies confirm the independent contribution of each component. Occluded training reduces FID from 113.1 to 39.9, eliminating template dependency while enabling occlusion-aware de-identification. |