Hybrid approach to efficient text extraction in complex color images

Texture-based methods and connected component (CC) methods have been widely used for text localization. However, these two primary methods have their own strength and weakness. This paper proposes a hybrid approach of the two methods for text localization in complex images. An automatically constructed MLP-based texture classifier can increase the recall rates for complex images with much less user intervention and no explicit feature extraction. The CC-based filtering based on the geometry and shape information enhances the precision rates without affecting overall performance. Then, the time-consuming texture analysis for less relevant pixels is avoided by using CAMShift. Our experimentation shows that the proposed hybrid approach leads to not only robust but also efficient text localization.