HSTENet: A deep learning-based scene text recognition system

In many computer vision applications such as intelligent transportation systems, autonomous navigation, digital content retrieval, and multilingual translation, detection and recognition of scenes play a vital role. These systems are fairly effective at document-oriented applications but less effective in real-world applications where an increased number of languages, orientation variation, noise, and distortion are present. To overcome these shortcomings, we are introducing a Hybrid Structured Text Extraction Network (HSTENet) for robust and accurate scene-text recognition that combines deep learning and machine learning. HSTENet features advanced modules such as edge detection with image preprocessing, support vector machine classification for detecting the textual part, and a line opening algorithm for fast segmentation of the image into regions. The recognition part relies on deep architecture, convolutional neural networks, attention-based connectionist temporal classification decoding, and bidirectional long short-term memory networks to recognize the high-precision target text. HSTENet was evaluated on three benchmark datasets—the International Conference on Document Analysis and Recognition, Common Objects in Context for Text, and a self-designed high-noise multilingual dataset—and obtained an F1-score of 91.9% and recognition accuracy of 93.8%, significantly outperforming other state-of-the-art models, including DeepText, convolutional recurrent neural networks, and connectionist text proposal networks. The framework is also polarization-invariant, real-time compatible (inference time 0.69 s), and deployable on edge devices. The results showed that HSTENet is highly efficient and automated for textual scene recognition, which could be applied to systems embedded in mobile platforms, self-contained platforms, and intelligent human–computer interaction.