具有仿射变换锚点的文字检测方法

Text Detection via Affine Transformation Anchor Proposals

针对现实场景图片中的文字区域具有仿射变换和多方向的问题,设计了一种带有仿射变换锚点,能够生成仿射变换文本预测框的文本检测网络。按照常见文字实例的仿真变换形状,给定了6种固定角度和5个固定变换量。检测过程中对预测候选框的角度和偏移值进行了调整拟合,最后对边框进行了回归,让检测结果更适应真实文字区域的边界框。与以往的文字检测网络相比,该检测方法能够有效适应文字区域,在检测精度和平均指标上分别有了7%和10%的提升。

Aiming at the problem of affine transformation and multi-direction in the text area in real scene pictures,a text detection network with affine transformation anchors is designed,which can generate affine transformation shape text prediction box.6 angles and 5 transformation amounts are given by the affine transformation amount of common text annotation.In the detection process,the angle and offset value of the predicted candidate box are adjusted and fitted,and finally the frame is regressed to make the detection result more suitable for the bounding box of the real text area.Compared with the previous text detection network,this detection method can effectively adapt to the text area,and the detection accuracy and average index have been improved by 7%and 10%respectively.