Lightweight Fully Convolutional Network for License Plate Detection

This paper presents an efficient lightweight Fully Convolutional Network(FCN) for license plate detection from complex scenes. Our network downscales input images for substantially accelerating proceeding and reducing the computational cost. Dense connections and dilated convolutions are adopted for combing multilevel and multiscale vision features. A fusion loss structure is appended during training to further improve prediction accuracy. For performance evaluation, we use a dataset consisting of 3977 images captured from diverse scenes under different conditions and widely used Caltech license plate dataset. Experiments show that the proposed method outperforms over many existing state-of-the-art methods and achieves a good tradeoff between high accuracy and computational costs.