End to End Text Detection and Recognition of Natural Images using Dictionary

In recent years, text detection and recognition in natural images have been extensively studied.In this study, a robust multi-oriented scene text localization system was proposed to obtain high efficiency in text detection based on a convolutional neural network (CNN). The proposed method includes three layers of feature extraction, feature-merging, and output. An improved ReLU layer (i.ReLU) is introduced in the feature extraction layer. An improved inception layer (i.inception) is also provided to detect texts with valuable information.An extra layer has been used to improve the feature extraction, which enables the proposed structureto detect multi-oriented even curved and vertical texts. We have proposed a pipeline framework for character recognition.The proposed pipeline framework consists of two parallel pipelines that are processed at the same time, and can recognize 62 characters. The first pipeline consists of cropped words and the second pipeline consists of text angles. Then, we formed a dictionary and used it to correct the possible error of the recognized words. Experiments on the ICDAR 2013, ICDAR 2015 and ICDAR 2019 datasets demonstrated the architectural superiority of the proposed structure over the previous works.