Abstract
One of the most computationally intensive parts in modern recognition systemsis an inference of deep neural networks that are used for image classification,segmentation, enhancement, and recognition. The growing popularity of edgecomputing makes us look for ways to reduce its time for mobile and embeddeddevices. One way to decrease the neural network inference time is to modify aneuron model to make it moreefficient for computations on a specific device.The example ofsuch a model is a bipolar morphological neuron model. The bipolarmorphological neuron is based on the idea of replacing multiplication withaddition and maximum operations. This model has been demonstrated for simpleimage classification with LeNet-like architectures [1]. In the paper, weintroduce a bipolar morphological ResNet (BM-ResNet) model obtained from a muchmore complex ResNet architecture by converting its layers to bipolarmorphological ones. We apply BM-ResNet to image classification on MNIST andCIFAR-10 datasets with only a moderate accuracy decrease from 99.3% to 99.1%and from 85.3% to 85.1%. We also estimate the computational complexity of theresulting model. We show that for the majority of ResNet layers, the consideredmodel requires 2.1-2.9 times fewer logic gates for implementation and 15-30%lower latency.