Generating Text Sequence Images for Recognition

Abstract

Recently, methods based on deep learning have dominated the field of text recognition. With a large number of training data, most of them can achieve the state-of-the-art performances. However, it is hard to harvest and label sufficient text sequence images from the real scenes. To mitigate this issue, several methods to synthesize text sequence images were proposed, yet they usually need complicated preceding or followup steps. In this work, we present a method which is able to generate infinite training data without any auxiliary pre/postprocess. We tackle the generation task as an image-to-image translation one and utilize conditional adversarial networks to produce realistic text sequence images in the light of the semantic ones. Some evaluation metrics are involved to assess our method and the results demonstrate that the caliber of the data is satisfactory. The code and dataset will be publicly available soon.

Realistic Text Replacement With Non-Uniform Style Conditioning

“…In [25], proxy metrics are trained to assess the model quality. The authors of [25] train a model on 8 million synthetically generated images. In our case, because of the size of our network, such validation would be unfeasible.…”

“…In [6], proxy metrics are trained to assess the model quality. The authors of [6] train a model on 8 million synthetically generated images.…”

“…Another work related to image generation and text is [25]. Its main focus is the augmentation of datasets for text recognition.…”

“…Its main focus is the augmentation of datasets for text recognition. The authors of [25] propose an adaptation of the Pix2Pix network to realistically color text images. The proposed architecture is composed of a combination of two Pix2Pix generators.…”

“…Unlike earlier works [1], [25], we design an architecture capable of text replacement of several text areas in one forward pass. To accomplish this, we introduce a novel nonuniform conditional normalization layer.…”