Today’s availability of enormous amounts of computational power in the form of relatively cheap GPUs and publicly accessible cloud computing facilities makes the training of large deep neural networks practical. Also embedded systems have been gaining in computational power and reducing their prices, making deployment of bigger neural networks on embedded systems feasible. In the scope of this diploma thesis the necessary components and methods for the implementation of deep neural networks for image classification trained on cloud computers and deployed on embedded systems are brought together and shown working. Several embedded systems were used with different neural network architectures and their capabilities, performance, resource usage, price and practicality compared. This serves as a guide to implement state of the art image classification on easily available and low cost embedded systems.
|
