Efficient Implementation of 2D Convolutions in Spatial Domain

Rad sadrži teoretsku razradu vezanu za matematički model konvolucije, terotiju vezanu za konvolucijske neuronske mreže s naglaskom na konvolucijski sloj, vrste konvolucijskih filtera, vrste konvolucije, relevantne konvolucijske algoritme. Rad sadrži implementacije naivnog algoritma, algoritma matričnog množenja, te implementaciju korištenjem biblioteke cuDNN, s naglaskom na brzine izvođenja programa. Korištena je i biblioteka OpenCV zbog prilagođenosti obradi slike.The paper contains theoretical processing related to the mathematical model of convolution, terotia associated with convolutional neural networks with emphasis on the convolutional layer, types of convolutional filters, types of convolutions, relevant convolution algorithms. The paper includes the implementation of a naive algorithm, matrix multiplication algorithm, and deployment of the cuDNN library, with an emphasis on program execution speeds. The OpenCV library was also used for custom image processing