Fast and lightweight text line detection on historical documents

Abstract

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.