Textareas (TA) detection in images with complex inhomogeneous background is complicated by the fact that the texts in these images are often not separated from other information clearly, and text is a part of this information. Despite the abundance of publications, currently there is no method and system of text areas detection, guaranteed to solve this problem with an acceptable quality. The article describes the developed information technology of text localization in images with complex background. The suggested by the authors' method of text areas localization in images with complex background, based on convolutional neural network, is a basis of proposed information technology. The method takes into account the image's multi-scale decomposition on wavelet basis and training on basis of a characters probabilistic model.
Using models generated in forming the training samples decreases dependence of the localization quality of the number and characteristics of real-world images containing TA, increases flexibility to create the training set, increases localization accuracy after training convolution neural network.
Performance testing of proposed information technology of text localization was conducted. Localization accuracy then on the image of the training sample was 99.9% and in the control sample - 87.7%.