Adaptation of Convolution and Batch Normalization Layer for CNN Implementation on FPGA

The article presents integration process of convolution and batch normalization layer for further implementation on FPGA. The convolution kernel is binarized and merged with batch normalization into a core and implemented on single DSP. The concept is proven on custom binarized convolutional neural network (CNN) that is trained in Matlab to solve object localization task. 16 b precision gives 1.3 % error on the output of joined convolution and batch normalization core. The localization accuracy decreases in average by 7 % from 74 % to 67 %, and it is still tolerable in embedded systems applications.