Uygur Words Localization in Single Frame Image of Complex Background

This paper proposes a localization method of Uygur words in video. First, the video color frame image is preprocessed, while the wavelet transformation is applied and texture entropy is calculated using three-stage wavelet in high frequency details. The entropy is regarded as an important parameter to extract texture characteristic of the image. Further segment each image into some sub-domains and calculate their statistical features. Then we determine the attribute characteristics of the image domain using neural network classifier, whereby the localization of word area is accomplished. Experimental results indicate that compared with the existing algorithms, this algorithm still has good effect on localization of Uygur words whose script, size and position are all uncertain, even under a complex background.