**Título:** Complexidade de Amostra para Projetar Operadores para Imagens Binárias pela Aprendizagem de Máquina

**Autores:** Kim, Hae Yong

**Resumo:** Binary operators (or filters) have a broad range of pratical applications. Several works have shown that binary operators can be sucessfully designed using a system based on the machine learning. Designing a binary operator by hand is a hard and annoying task. The use of a learning system allows the user to specify operators easily, by simply feeding the system with pairs of in-out sample images. In this paper, we present a set of techniques to estimate the sample complexity of the binary operator learning problem. The sample complexity is the quantity of training samples needed to get, with probability at least (1-δ), an operator with an error rate at most ε. We make use of the PAC (Probably Approximately Correct) learning theory to compute it, to both noise-free and noisy cases. As the PAC theory usually overestimates the sample complexity, the statistical estimation is used to calculate a posteriori, a tight error rate. We also show how the minimal error rate for a given noisy problem can be estimated. Finally, we apply the theory developed to analyze the sample complexity and the error rate for the spatial resolution increasing of electronic documents with printed characters by machine learning.

**Palavras-chave:**

**Páginas:** 6

**Código DOI:** 10.21528/CBRN2001-030

**Artigo em pdf:** 5cbrn_030.pdf

**Arquivo BibTex:** 5cbrn_030.bib