End-to-End Vessel Plate Number Detection and Recognition Using Deep Convolutional Neural Networks and LSTMs

This paper proposes a new framework for vessel plate number detection and recognition in natural scene images. There are a lot of existing methods which treat vessel plate number detection and identification as two separate tasks. This paper presents a unified end-to-end trainable deep network, which can simultaneously locate and recognize the vessel plate number. The computational and visual information have been shared between detection and recognition. In addition, the traditional text detection methods use horizontal bounding box proposals, which cannot adapt to text regions in real-world applications. Therefore, this paper use the arbitrary-oriented proposals instead of the horizontal proposals. The experiments on the vessel image dataset show that our method is feasible and efficient in the vessel plate number detection and recognition.