Object classification using neural networks with binary input and binary feature extraction

Sergey Poslavskiy, Dmitriy Shashev, Stanislav Shidlovskiy
Although machine learning by its nature, has been resource-intensive, mul- tiple resource efficient alternative approaches have been made in the field of embedded systems. Researchers over the years, have specially shown interest and thus it has fueled the research output, in the field of autonomous naviga- tion, IoT and distributed embedded systems. These approaches are aimed at finding a trade-off between performance and resource consumption in terms of computational costs and energy efficiency. The development of appropriate algorithms is one of the main tasks of modern research in the field of machine learning and the key to ensuring smooth adaptation of machine learning tech- nologies in an environment with limited or distributed computing resources. These algorithms can be divided into four non-mutually exclusive categories: quantization of neural networks, network pruning, structural efficient algorithms, binarized neural networks. Our approach is proposed for working with binary images and extracting binary features for training neural networks This approach provides the possibility of an implementation based on the architecture of a reconfigurable computing systems. Our approach was used to build a neural network architecture. The resulting architecture was trained and tested on the MNIST dataset. The results were compared with the results of the LeNet-5 architecture.