Scene Text Detection Using Attention with Depthwise Separable Convolutions

In spite of significant research efforts, the existing scene text detection methods fall short of the challenges and requirements posed in real-life applications. In natural scenes, text segments exhibit a wide range of shape complexities, scale, and font property variations, and they appear mostly incidental. Furthermore, the computational requirement of the detector is an important factor for real-time operation. To address the aforementioned issues, the paper presents a novel scene text detector using a deep convolutional network which efficiently detects arbitrary oriented and complex-shaped text segments from natural scenes and predicts quadrilateral bounding boxes around text segments. The proposed network is designed in a U-shape architecture with the careful incorporation of skip connections to capture complex text attributes at multiple scales. For addressing the computational requirement of the input processing, the proposed scene text detector uses the MobileNet model as the backbone that is designed on depthwise separable convolutions. The network design is integrated with text attention blocks to enhance the learning ability of our detector, where the attention blocks are based on efficient channel attention. The network is trained in a multi-objective formulation supported by a novel text-aware non-maximal procedure to generate final text bounding box predictions. On extensive evaluations on ICDAR2013, ICDAR2015, MSRA-TD500, and COCOText datasets, the paper reports detection F-scores of 0.910, 0.879, 0.830, and 0.617, respectively.

Automated summary

Despite significant research efforts, existing scene text detection methods are insufficient for real-life applications due to challenges posed by complex shapes, scale variations, and font properties in text segments. This paper proposes a novel scene text detector using a deep convolutional network that efficiently detects arbitrary oriented and complex-shaped text segments from natural scenes. The network design incorporates skip connections and text attention blocks based on efficient channel attention to capture complex text attributes at multiple scales. Extensive evaluations on various datasets show high detection F-scores.