Text Detection And Recognition Based On Component Tree And Hough Forest

Text detection and recognition in natural scenes play an important role in image understanding, including image annotation, video surveillance analysis, location-based services and instant translation, etc. Due to the complexity of natural scenes, as well as the diversity of the text itself, there is no perfect solution for this problem. In most of current system design, detection and recognition are isolated and processed separately, which recognition is premised on accurately detection and text characteristics are not fully utilized under detection.Two effective solutions are presented. The first option is a method based on unsupervised feature learned by deep convolution autoencoder network, as well as mixed feature for different scales by sparse coding. The detection result shows good, but the solution cannot meet realtime demand for the large amount of calculation. Another option, a unified framework for detection and recognition based on multi-class Hough forest is proposed. The scheme combines detection and recognition process in order to decrease the computation complexity and improve detection accuracy. For the sake of improving the performance when the quantity of classes increases, as well as improve accuracy with uncertain scale, component tree, which enhances the constraint of Maximally Stable External Regions, is used for extracting connected component with hierarchy, while a set of features based on text characteristics is extracted and feeds to a classifier. With the help of the classifier, the scale of the target is determined and all candidate texts are located, which builds the foundation of subsequent stage for fine positioning and recognition.The main contribution of this paper is as follows:1. Apply unsupervised feature, which learned by deep convolutional network under different scale and mixed by sparse coding, to text detection in natural scenes;2. Through modify the split function for Hough forest, a unified framework for text detection and recognition based on multi-class Hough forest classifier is proposed;3. The proposed use of the component tree, combined with a set of features designed for extracting text candidate regions, as well as to determine the scale of the target.Experiments show that the scheme is competitive with current optimal solutions in both detection and recognition.