In this project, we use the cats & dogs database:

https://github.com/sankit1/cv-tricks.com/tree/master/Tensorflow-tutorials/tutorial-2-image-classifier

TRAINING PHASE:

In training process, we use a DELL Workstation. And, the Ubuntu 14.02, the tensorflow with GPU and the Cuda, etc.  were installed on the computer.

The training code was modified to save the weights of each layers at the convolution neural network as text files.

 

TESTING PHASE in a TARGET SYSTEM WITHOUT INSTALLING TENSORFLOW (The RaspberryPi, the OrangePi, the NanoPi, etc.):

The convolution neural network (CNN) model is writen in C language on the target systems without installing the tensorflow. These text files obtained in the training phase were used to determine all weights of the CNN. We use the python language as a glue. The Python forms the framework, and call the CNN as import C file. The python obtains images from the camera of the computer at real-time by using the opencv library. At below 1 second, the Python framework made a decision by using the camera and import CNN file.

The Following Works were done for these purposes:

1)    RGB-images (with two-dimensional) were transfered from the python to C, or from the C to the Python.

 

2)    The convolution neural network (CNN) model is writen in C language. The program written in C was shared among the cores of a Embedded Linux System by using the omp library to increase the speed. The CNN program written in C was compiled on the Embedded Linux System (The RaspberryPi, the OrangePi, the NanoPi, etc.). The file with (.so) extansion was called from the python as a import file.