Faster glare detection on document images

Abstract

Glare on images is caused by a bright light source or its reflection. This effect can hide or corrupt important information on a document image, negatively affecting the overall quality of text recognition and data extraction. In this paper, we present a new method for glare detection which is based on our prior work on barcode detection, where we used a convolutional neural network (CNN) with consecutive downsampling and context aggregation modules to extract relevant features. In this paper, we propose a similar CNN to create a segmentation map for image glare. The proposed model is fast, lightweight, and outperforms previous approaches in both inference time and quality. We also introduce a new dataset of 687 document images with glare. The glare regions on the document images were marked and the dataset itself has been made publicly available. All of the measurements and comparisons discussed in this paper were performed using this dataset. We managed to obtain an F1-measure of 0.812 with an average run time of 24.3 ms on an iPhone XS. The experimental data is available at https://github.com/RodinDmitry/glare.