Convolution Neural Network Based Deep Features for Text Recognition in Multi-Type Images

Attaining worthy recognition rate for text in multi-type images is quite a challenging. In this work, we propose a model which is able to recognize the text in multi-type images consists of scene, video and born-digital images with better recognition rate. Initially, a wavelet transformation based sliding window approach is used to extract the high frequency band for every sliding window. Subsequently, K-means clustering is used to extract the text candidates by reducing the background complexity and noise. In addition to this, the direction of each window is detected through PCA. Secondly, a four-layer convolutional neural network based model is designed to recognize the text in multi-type images. An extensive experimentation is conducted on the ICDAR13 scene, ICDAR15 video and ICDAR11 born-digital word images, demonstrated that proposed method outperforms the other existing methods.