A deep learning approach for robust, multi-oriented, and curved text detection

Automatic text localization and segmentation in a normal environment with vertical or curved texts are core elements of numerous tasks comprising the identification of vehicles and self-driving cars, and preparing significant information from real scenes to visually impaired people. Nevertheless, texts in the real environment can be discovered with a high level of angles, profiles, dimensions, and colors which is an arduous process to detect. In this paper, a new framework based on a convolutional neural network (CNN) is introduced to obtain high efficiency in detecting text even in the presence of a complex background. Due to using a new inception layer and an improved ReLU layer, an excellent result is gained to detect text even in the presence of complex backgrounds. At first, four new m.ReLU layers are employed to explore low-level visual features. The new m.ReLU building block and inception layer are optimized to detect vital information maximally. The effect of stacking up inception layers (kernels with the dimension of 3 × 3 or bigger) is explored and it is demonstrated that this strategy is capable of obtaining mostly varying-sized texts further successfully than a linear chain of convolution layers (Conv layers). The suggested text detection algorithm is conducted in four well-known databases, namely ICDAR 2013, ICDAR 2015, ICDAR 2017, and ICDAR 2019. Text detection results on all mentioned databases with the highest recall of 94.2%, precision of 95.6%, and F-score of 94.8% illustrate that the developed strategy outperforms the state-of-the-art frameworks.