Experimental evaluation of recognizing digits and letters by convolutional neural networks

U radu su objašnjeni osnovni pojmovi potrebni za razumijevanje umjetne inteligencije, neuronskih mreža te konvolucijskih neuronskih mreža. Njihova glavna karakteristika je učenje na danom skupu uzoraka za razliku od ostalih analitičkih metoda rješavanja problema. Predmet rada je ,osim učenja neuronskih mreža, sakupljanje uzoraka koji se dovode na ulaz mreže. Korištena je Javascript biblioteka DrawMe za označavanje markica s brojevima te su razvijene funkcije za rotaciju i izrezivanje markica i brojeva. Za treniranje konvolucijske neuronske mreže koristi gotova biblioteka Caffe.The paper explains the basic concepts necessary to understand artificial intelligence, neural networks and convolution neural networks. Their main caracteristic is learning on the given set of samples, unlike other analytical methods to solve problems. The subject of the paper is, besides learning neural networks, collecting samples which are being brought on network entrance. Javascript library DrawMe was used for tagging stamps with numbers and addition functions were made for rotation and cropping marks and numbers have been developed. For training on convolution neural networks is used library Caffe