[论文解读]|PAN:Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network

题目:Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network
作者:Wenhai Wang, Enze Xie,Xiaoge Song, Yuhang Zang, Wenjia Wang, Tong Luy, Gang Yu, and Chunhua Shen

Motivation:
作者认为目前场景文字检测难点在于速度和效果间的差异,及变化形状的文字检测。作者在考虑速度与效果兼得的前提下提出PAN模型。

Method
[论文解读]|PAN:Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network
作者提出如上图的pipeline的PAN方法。
[论文解读]|PAN:Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network
PAN模型包括FPEM和FFM网络
**FPEM:**是一个可变大小的计算速度快的网络,以获得大小特征图的信息。
[论文解读]|PAN:Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network
**FFM(特征整合模型)????*对上面多个FPEM网络进行融合
[论文解读]|PAN:Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network
最后通过上采样将它们Concatenate到一起。

模型最后预测三种信息:
1、文字区域
2、文字kernel
3、文字kernel的相似向量

Pixel Aggregation:
由于文字与文字可能出现部分重叠,作者采用Kernel的方法。使文字的像素与kernel的距离最小化,其Loss为:
[论文解读]|PAN:Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network
其中,[论文解读]|PAN:Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network

然而文字与文字间的kernel要保持一定的距离以区分开来,因此,作者又使用:
[论文解读]|PAN:Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network
LossFunction:
[论文解读]|PAN:Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network
其中文字区域和kernel预测loss为:
[论文解读]|PAN:Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network
Experiment:
不多说,直接上图。
[论文解读]|PAN:Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network
[论文解读]|PAN:Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network
[论文解读]|PAN:Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network
[论文解读]|PAN:Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network
作者在多个数据集上做过实验,而且效果都不错,作者验证很充分。

END