关键词: 文本检测, 自然场景, 类别失衡, 边界框校准

Abstract:

A text detection method based on deep learning is proposed for multi-directional text objects in natural scenes. When designing the anchor, the directional feature of the anchor is removed but the aspect ratio feature is preserved. When covering the same aspect ratio range, the number of anchors is reduced, thereby alleviating the influence of the imbalance of positive and negative samples in dense sampling. In  addition, in the post-processing stage of the method, a bounding box calibration algorithm is proposed, which uses the Maximally Stable Extremal Region（MSER） to obtain the character edge information, and then shrinks or expands the bounding box through rule-based logic judgment, thereby achieving the purpose of  bounding box calibration. The effectiveness of the proposed bounding box calibration algorithm is verified by testing and comparison on the public dataset ICDAR2015.

Key words: text detection, natural scene, category imbalance, bounding box calibration