Text Sign Detection Method based on Dual Feature Fusion

Abstract

Text sign detection in natural scenes faces the challenge of complex background and various colors, which imposes a great burden on irregular text recognition. Therefore, special classification and detection methods are needed. This paper proposes a text sign detection method based on dual feature fusion, which improves the detection accuracy by fusing contour and color features.  In the aspect of contour features, the spatial pyramid matching model is combined with SIFT feature extraction technology, and the self-growing neural network is used to adaptively determine the number of feature categories. In terms of color processing, the HSV color space is improved. By quantizing the hue and saturation components and sorting the color distribution, the color features that are robust to illumination changes are obtained.  The most important innovation is to concatenate the contour histogram and the improved color histogram to form a comprehensive feature representation with dual feature fusion. The experimental results show that the proposed method performs well in the test containing thousands of street view images, with an accuracy of 89.16% for positive samples and 94.3% for negative samples, which is significantly better than the method using a single feature.  This result verifies that considering both shape and color information can better recognize text signs, which provides an effective technical solution for practical applications.