Deep learning served for feature engineering stuff (experiments).
For this ID processing, tested on more images (random pictured) with 96.8% accuracy (2 failed on 63 images in total).
To transplant to mobile end, the potential way might be:
1 To use smaller structure;
2 To use lite frameworks on, such as tensorflow lite, ncnn, etc., which supports Android and IOS, and they all have C++ API.
Note:
OpenCV 3 has introduced cv::dnn module, but after initial try there is still no luck to load a freezen tf graph, and throw the error as "Too much unspecific tensor..." or something.
Solution:
1 Get rid of these tensor nodes such as dropout or batch-normalization related stuff, using data pre-processing to replace those functions for generalization;
2 Introducing other framework API in end, caffe2 is a good try, some of them give the pre-trained MobileNet or YOLO stuff, and start with these model and transfer can be great tradeoff;
3 Keep patient until opencv or tensorfolw rocks.