License Plate Detection via Information Maximization

License plate (LP) detection in the wild remains challenging due to diversity of environmental conditions. Nevertheless, prior solutions have focused on controlled environments, such as when LP images frequently emerge from an approximately frontal viewpoint and without scene text which might be mistaken for LP. However, even state-of-the-art object detectors, their performance is not satisfactory real-world suffering various types degradation. To solve these problems, we propose a novel end-to-end framework robust detection, designed settings. Our contribution threefold: (1) A information-theoretic learning that takes advantage shared encoder, detector (excluding LP) simultaneously; (2) Localization refinement generalizing bounding box regression network complement ambiguous results; (3) large-scale, comprehensive dataset, LPST-110K, representing unconstrained scenes including annotations. Computational tests show proposed model outperforms other methods variety datasets.