End-to-End Multi-line License Plate Recognition with Cascaded Perception

Due to the irregular layout, multi-line license plates are challenging to recognize, and previous methods cannot recognize them effectively and efficiently. In this work, we propose an end-to-end multi-line license plate recognition network, which cascades global type perception and parallel character perception to enhance recognition performance and inference speed. Specifically, we first utilize self-information mining to extract global features to perceive plate type and character layout, improving recognition performance. Then, we use the reading order to attend plate characters parallelly, strengthening inference speed. Finally, we propose extracting recognition features from shallow layers of the backbone to construct an end-to-end detection and recognition network. This way, it can reduce error accumulation and retain more plate information, such as character stroke and layout, to enhance recognition. Experiments on three datasets prove our method can achieve state-of-the-art recognition performance, and cross-dataset experiments on two datasets verify the generality of our method. Moreover, our method can achieve a breakneck inference speed of 104 FPS with a small backbone while outperforming most comparative methods in recognition.