(A) real-time text detection network via convolution decomposition in natural scene images

Abstract

Text detection has received considerable attention because of its high usability. Recently proposed methods exhibit powerful detection performance using deep networks, requiring a complex computation process that is becoming an obstacle to commercializing the technology. The proposed method aims to localize text in real-time while maintaining competitive detection performance. To this end, we decompose a typical $M \times M$ convolution kernel into smaller kernels. We introduce a depthwise cross convolution method for convolution decomposition. By doing this, the proposed method can process 40 images per second. It is three times as fast as the previous fastest deep-network-based method. The proposed method not only localizes text, but also predicts text line orientation. This allows the proposed method to increase utilization because text is not always horizontally aligned.