The note about the original paper: SSD: Single Shot MultiBox Detector can be found here.
This practice is inspired by ssd-plate_detection, thanks to solace_hyh.
The detail of the above code can read my blog: http://blog.csdn.net/u010167269/article/details/52851667.
Meanwhile, I have uploaded my training caffemodel to WeiYun. The link：http://share.weiyun.com/1c544de66be06ea04774fd11e820a780, the extraction code：GjjwTB
Some examples of the scene text detection:
