Tiny CNN for feature point description for document analysis: approach and dataset

Abstract

In this paper, we study the problem of feature points description in the context of document analysis and template matching. Our study shows that specific training data is required for the task especially if we are to train a lightweight neural network that will be usable on devices with limited computational resources. In this paper, we construct and provide a dataset of photo and synthetically generated images and a method of training patches generation from it. We prove the effectiveness of this data by training a lightweight neural network and show how it performs in both general and documents patches matching. The training was done on the provided dataset in comparison with HPatches training dataset and for the testing, we solve HPatches testing framework tasks and template matching task on two publicly available datasets with various documents pictured on complex backgrounds: MIDV-500 and MIDV-2019.

Training of binary neural network models using continuous approximation

“…Deep neural networks successfully perform many tasks, ranging from video recognition to text generation or dialog systems [1][2][3]. The main trajectory of developing deep neural network models is the creation of new architectures and enhancing the existing ones to improve accuracy and solve new problems [4][5][6]. However, modern deep neural networks contain hundreds of layers and millions of trainable parameters.…”