GMIF: A Gated Multiscale Input Feature Fusion Scheme for Scene Text Detection

The feature fusion of the multi-scale features plays a significant role in localizing text instances of different sizes in the scene text detection (STD) paradigm. The existing approaches are not sufficient to tackle the issues of multi-scale text; consequently, their performance also varies with the text size. Here, we propose a gated multi-scale input feature fusion (GMIF) approach to overcome this issue in STD. The GMIF generates the local features from down-scaled input images and propagates these features from low resolution to the higher resolution global features through a gated recurrent unit-like mechanism. The consistent performance of the GMIF is validated with different text instance sizes of the test-set of the Total-text dataset. The GMIF obtained the performance in range (Precision 88.554-89.106, Recall 85.452-85.790, and f-measures 87.072 - 87.417) with marginal deviation, whereas the current state-of-the-art method, DBNet++, acquired in range (Precision 73.005-82.666, Recall 80.912-87.274, and f-measures 76.755 - 84.183) with significant deviation. Besides this, GMIF also achieved the best performance (f-measures) over ICDAR 2015 (as 88.0), Total-Text (as 87.4), and the second-best over the MSRA-TD500 (as 85.2) dataset. We have conducted an ablation study to show the impact of different components of the GMIF on the STD tasks, which shows the effectiveness of the overall GMIF approach.