Made available in DSpace on 2016-08-17T14:52:43Z (GMT). No. of bitstreams: 1 Andre Borges Cavalcante.pdf: 1739525 bytes, checksum: 2073615c7df203b086d5c76276905a35 (MD5) Previous issue date: 2008-02-25 / Efficient coding of natural images yields filters similar to the Gabor-like receptive fields of simple cells of primary visual cortex. However, natural and man-made images have different statistical proprieties. Here we show that a simple theoretical analysis of power spectra in a sparse model suggests that natural and man-made images would need specific filters for each group. Indeed, when applying sparse coding to man-made scenes, we found both Gabor and Haar wavelet-like filters. Furthermore, we found that man-made images when projected on those filters yielded smaller mean squared error than when projected on Gabor-like filters only. Thus, as natural and man-made images require different filters to be efficiently represented, these results suggest that besides Gabor, the primary visual cortex should also have cells with Haar-like receptive fields. / A codificação eficiente de imagens naturais gera filtros similares às wavelets de Gabor que relembram os campos receptivos de células simples do córtex visual primário. No entanto, imagens naturais e urbanas tem características estatísticas diferentes. Será mostrado que uma simples análise do espectro de potência em um modelo eficiente sugere que imagens naturais e urbanas requerem filtros específicos para cada grupo. De fato, aplicando codificação eficiente à imagens urbanas, encontramos filtros similares às wavelets de Gabor e de Haar. Além disso, observou-se que imagens urbanas quando projetadas nesses filtros geraram um menor erro médio quadrático do que quando projetadas somente em filtros de similares a Gabor. Desta forma, como imagens naturais e urbanas requerem filtros diferentes para serem representadas de forma eficiente, estes resultados sugerem que além de Gabor, o córtex visual primário também deve possuir células com campos receptivos similares às wavelets de Haar.

A major challenge in training machine learning models is the gathering and labeling of a sufficiently large training data set. A common solution is the use of synthetically generated data set to expand or replace a real data set. This paper examines the performance of a machine learning model trained on synthetic data set versus the same model trained on real data. This approach was applied to the problem of character recognition using a machine learning model that implements convolutional neural networks. A synthetic data set of 1’240’000 images and two real data sets, Char74k and ICDAR 2003, were used. The result was that the model trained on the synthetic data set achieved an accuracy that was about 50% better than the accuracy of the same model trained on the real data set. / Vid utvecklandet av maskininlärningsmodeller kan avsaknaden av ett tillräckligt stort dataset för träning utgöra ett problem. En vanlig lösning är att använda syntetiskt genererad data för att antingen utöka eller helt ersätta ett dataset med verklig data. Denna uppsats undersöker prestationen av en maskininlärningsmodell tränad på syntetisk data jämfört med samma modell tränad på verklig data. Detta applicerades på problemet att använda ett konvolutionärt neuralt nätverk för att tyda tecken i bilder från ”naturliga” miljöer. Ett syntetiskt dataset bestående av 1’240’000 samt två stycken dataset med tecken från bilder, Char74K och ICDAR2003, användes. Resultatet visar att en modell tränad på det syntetiska datasetet presterade ca 50% bättre än samma modell tränad på Char74K.