Abstract:

 For character preclassification, a generalized confidence threshold determination rule and multi-resolution histograms of oriented gradients feature are proposed. It is shown that an optimal tradeoff between “error rate”and “average candidate number” can be obtained based on the proposed rule. From the Gaussian pyramid of a gradient image,multi-resolution histograms of oriented gradients are extracted. Experimental results indicate superior performance of the proposed rule and features. 

Key words: pre-classification; generalized confidence; multi-resolution histogram of oriented gradient