The Research Of Text Detection Method In Trademark Images

The image of trademark generally consists of some patterns and texts.With the continuous development of commodity economy and the continuous improvement of people's awareness of intellectual property rights,the trademarks play an increasingly important role in the process of commodity circulation.The number of trademark applications is increasing,and the trademark infringement cases,especially those related to trademark words,are also increasing.The relevant work of the competent departments needs the assistance of automation means,and the text detectors are very important as the first step of some systems.However,due to the complex structure of the trademark texts and other problems,the existing text detection methods in the detection of trademark text may cause the situation of wrong results and missing results.Inspired by the application of some text detection methods based on deep learning,this thesis analyzes some text detection algorithms based on deep learning,we decide to design a text detection model firstly,and then train our model by a semi supervised training strategy,so that the model can detect characters,finally,the characters are grouped into thetext lines.The trademark character detection model improves the robustness of the network for text with different sizes through the multi-layer feature fusion structure in the network.Because there is no public trademark character data set,and the production cost of the character data set is high,the trademark text detection network proposed in this paper adopts semi supervised learning strategy.In order to facilitate the experiment,4500 trademark images were collected and the trademark text data set was established by manual annotation.The experimental results on manually annotated data sets show that the proposed method outputs better detect results in the complex structure and mixed Chinese and English trademark text detection task than other deep learning methods,especially the detection algorithm can detect text lines directly.