A Text Recognition Algorithm Based On Deep Learning Feature Fusion Using Preprocessing Technology

The image carries a variety of information,and the quantity of digital image data is great.Nowadays,image is an important source to obtain information,in order to quickly understand the image information is needed to assistance with the characters.The text in the images can reflect the meaning of the image,interpret the image content,save a lot of time for image understanding,and avoid the interference of redundant information.Therefore,the recognition of the text in the image is of great significance for quickly understand the image content.The text information extraction in natural scenes is challenged by the problem of interference information,complicated styles,and irregular arrangement,etc.The text recognition method based on CNN and LSTM feature fusion algorithm with preprocessing technology is proposed.First,PCA dimension reduction and RP random projection processing are performed on the image.Then,using CNN network to obtain high-level abstract visual features from the underlying image pixels,and using the local perceptual characteristics of the CNN network to establish the position relationship between the high-level features and the underlying pixels,and then using the LSTM network of the bidirectional RNN model to capture global features with short-term and long-term memory module.Finally,the classification of the Euclidean distance classifier is applied to obtain the recognition result.The effect of image feature dimension,CNN hidden layers,and preprocessing strategy on character recognition rate are studied.We got the feature dimension,hidden layers of CNN and the fusion strategy of preprocessing through experiment.On the ICAND image database,it is determined that the depth learning hidden layer is 17,the feature dimension is 45,the preprocessing fusion coefficient is 0.6:0.4,and the recognition rate is up to 97%.Comparative experiment is designed to study the PCA with Euclidean distance classification,random projection RP with Euclidean distance classification,convolutional neural network with Euclidean distance classification,and the method proposed in this paper.In the ICAND image database,under the condition of the hidden layer is 17,the feature dimension is45,and the preprocessing fusion coefficient is 0.6:0.4.It is concluded that the accurate rate of recognition is up to 97%,indicating that the proposed algorithm has a good ability of feature representation.The robustness verification experiment was designed on ICARD library,IIIT5 K library,and self-built library.Under the condition of the hidden layer is 17,the feature dimension is 45,and the preprocessing fusion coefficient is 0.6:0.4,the recognition rate of using Euclidean distance is about 94%.The experimental results shows that the paper algorithm has strong robustness bytesting in different image libraries.In order to verify the effectiveness of the algorithm,the preprocessing algorithm and the CNN algorithm are tested in the ICAND library with the condition of integrated graphics,stand-alone mode,and fixed parameters.The recognition rate of the preprocessing algorithm is98%,which is 44% higher than the CNN algorithm and cost less time.It is shows that the preprocessing algorithm is effective and improves the efficiency.In order to verify the anti-interference ability of the algorithm,the paper algorithm and the CNN algorithm are tested in the ICAND library with the condition of different light intensity,the occluded object,whether the font is complete or not.With different light intensity,the recognition rate of proposed algorithm is 98%,which is 44% higher than the CNN algorithm,and the running time is 132 s,which is shortened by 20 times.In the the condition of occluded object,the recognition rate of proposed algorithm is 98.05%,which is 50% higher than CNN algorithm,and running time is 64 s,which is shortened by 14 times;When the font is not regular,the recognition rate of proposed algorithm is 98.55%,which is 40% higher than CNN algorithm,and running time is 74 s,which is shortened by 16 times.It is verified that the proposed algorithm can maintain better detection performance under different interference environments and has better timeliness.