PROJECT TITLE :

Towards End-to-End Text Spotting in Natural Scenes

ABSTRACT:

The ability to recognize text within images of natural scenes is critical for a wide variety of image understanding tasks. Text detection and recognition are both subtasks that are included in it. In this work, we propose a unified network that simultaneously localizes and recognizes text with a single forward pass. By doing so, we eliminate the need for intermediate processes such as image cropping and feature re-calculation, word separation, and character grouping. The overall framework is trained from beginning to end, and it is able to recognize text in a variety of different shapes. The convolutional features are only ever computed once, and this information is then shared between the detection and recognition modules. Training on multiple tasks helps students become better at discriminating between similar features, which in turn improves their overall performance. Robust solutions for the problem of text irregularity have been developed through the utilization of a 2D attention model in word recognition. The attention model provides the spatial location for each character, which not only aids in the extraction of local features during word recognition but also indicates an orientation angle to refine text localization. This is accomplished through the attention model's ability to provide an orientation angle. Experiments have shown that our proposed method is capable of achieving state-of-the-art performance on a number of text spotting benchmarks that are commonly used. These benchmarks include both regular and irregular datasets. For the purpose of determining whether or not each module design is successful, exhaustive ablation experiments are carried out.