SegLink

PaperReading TextDetection 2017.11

1. main idea

• Decomposing text into segments and links.

• Some Words

  • aspect ratio
    the aspect ratio of a rectangle is the ratio of its longer side to its shorter side – the ratio of width to height。
    简单的说，就是宽高比。

2. Segment Linking

• CNN model

  • Architecture
    pretrained VGG-16
  • Segment Detection
  • Within-Layer Links
  • Cross-Layer Link Detection
    采用within-layer links的方法，同样的segment会在不同的feature map被检测到，会产生冗余信息。
    在这paper中，前一层的feature map的大小总是后一层的2倍，使得这个性质有效，要确保feature map的size总是偶数。
  • Output of a Convolutional Predictor

• Combining Segments with Links
  First,the output segments and links are filtered by their confidence score.
  Then,taking segments as nodes and links as edges,we construct a graph over them.
  DFS(Depth First Search).
  Finally, combine segments

  • 如何组合？
    求出平均旋转角度，利用最小二乘法拟合出中心点的直线，求出实际的中心点在这条线上的投影，最后找出宽高。注意坐标的变换。

3. Training

• Groundtruths of Sgements adn Links

  • Segments
    假设图片中只有一个groundtruth word, a default box is labeled as positive
    1) 中心点在word bounding box 之内
    2) box size 和 word height 之比最大值小于 $1.5$。
    考虑多个词的情况
    A default box is labeled as negative if it does not meet the above mentioned criteria.
    Otherwise, it is labeled as positive and matched to the word that has the closest size.
  • Links
    1) 连接的box都是positive
    2) 连接的box都连接到the same word

• Optimization

  • Objective
    3个loss的组合
    softmax loss + L1 loss
  • Online Hard Negative Mining
    正负样本不均衡
    采取策略保证pos:neg最大为3:1
  • Data Augmentation
    get the patch