Scene Text Detection Using Attention with Depthwise Separable Convolutions

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