Identification of Mongolian and Chinese Species in Natural Scenes Based on Convolutional Neural Network

摘要

Multilingual texts are very common in natural scenes. Dilterent languages have different shapes and structure. According to the language characteristics to choose the appropriate method can reduce the recognition error rate during text detection and recognition. This paper mainly proposes a lightweight convolutional neural network LTCNN for natural scene images mixed with Chinese characters and Mongolian. Firstly, collecting pictures contained both Chinese characters and Mongolian; then using perspective transformation to process the slanted text image, and using gamma transformation to enhance the image of underexposed or overexposed pictures;nest using the improved EAST model for the preprocessed pictures to extract text makes the data set needed for training the network; finally, the pictures in the data set are fed into the LTCNN network for training and identification of Mongolian and Chinese species. Experimental results show that the accuracy of Chinese and Mongolian recognition using this method can be close to 90% accuracy. Index Terms—language recognition,LTCNN,text correction.