We introduce a novel method for handwriting text line detection, which provides a balance between accuracy and computational efficiency and is suitable for mobile and embedded devices. We propose a lightweight convolutional neural network with only 22K trainable parameters which consists of two parts: downsampling module and context aggregation module (CAM). CAM uses stacked dilated convolutions to provide large receptive field for output features. The downsampling module allows a further reduction of latency. Our network has two output channels, one for baseline prediction and the other for detection of text lines without ascenders/descenders. This modification has a positive impact on accuracy. The proposed method achieves an F-measure of 0.873 on the ICDAR 2017 cBAD complex track with a network inference time of 8 ms on NVIDIA GeForce GTX 1080Ti card and another 12 ms for subsequent postprocessing operations on CPU, which is significantly faster than other existing methods with comparable accuracy.
CITATION STYLE
Melnikov, A., & Zagaynov, I. (2020). Fast and lightweight text line detection on historical documents. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 12116 LNCS, pp. 441–450). Springer. https://doi.org/10.1007/978-3-030-57058-3_31
Mendeley helps you to discover research relevant for your work.