As long as dlib NN api is not supported in Python, I wrote a simple python and C++ wrapper to send image into dlib C++ code from arbitrary container (I have int* for Python)
dlib_image_from_array.h does not need any dependency except dlib. Examples in the corresponding subfolders show usage in C++ and Python, I used OpenCV to quickly create different matrices. The code accepts int * data along with dimension, depth and channel data, but you can easily use any peculiar container by modifying variable types and navigate_in_array() function.
The code is based on Dlib's png_loader.h and does not support 16-bit image data except non-alpha grayscale. 16-bit pixel values are converted to 8-bit during input, see davisking/dlib#428
Use cmake scripts to compile. Compile shared library (.so) to import into Python module. Don't forget to add cmake . -DCMAKE_BUILD_TYPE=Release or dlib will work slow.
Tested on Dlib 19, OpenCV3, Python 2.7 and Python 3.5.
Upd: Recently I found this question http://stackoverflow.com/questions/41991651/how-to-assign-memory-images-to-a-dlib-array2d-or-image
so may be I will try to make better implementation sometime, based on these hints.