Natural Scene Mongolian Text Detection Based on Convolutional Neural Network and MSER

Abstract

Maximum Stable Extreme Region (MSER) is the most influential algorithm in text detection. However, due to the complex and varied background of Mongolian text in natural scene images, it is difficult to distinguish between text and non-text connected regions, thus reducing the robustness of the MSER algorithm. Therefore, this paper proposes to extract the connected regions in the natural scene pictures by applying MSER, and then uses the convolutional neural network (CNN) to train a high-performance text classifier to classify the extracted connected regions, and finally obtaining the final detection results. This paper evaluates the proposed method on the CSIMU-MTR dataset established by the School of Computer Science, Inner Mongolia University. The recall rate is 0.75, the accuracy rate is 0.83, and the F-score is 0.79, which is significantly higher than the previous method. It shows the effectiveness of the proposed Mongolian text detection method for natural scenes.