Text Recognition Of Relay Protection Equipment Image Based On Deep Learning

Relay protection equipment plays an important role in the stable operation of power system,which needs constant supervision and maintenance.The current manual inspection method can not meet the current growing scale of relay protection equipment,and the manually extracted equipment information is prone to errors,which may have a serious impact on the operation of the system.It can greatly improve the work efficiency and reduce the errors caused by manual negligence by manually taking the image of relay protection equipment and then sending it into the system to automatically identify the equipment information.With the breakthrough of deep learning in image recognition,this paper proposes a text recognition method based on deep learning according to the image properties of relay protection equipmentIn the part of text detection,a text detection algorithm based on the combination of improved mser algorithm and CNN(convolutional neural network)is proposed.According to the two limitations of mser algorithm,a multi-channel mser fusion algorithm based on light equalization and NMS algorithm are designed to filter the candidate text boxes;Then a text box classification network based on CNN is proposed.Firstly,on the basis of VGG network model,the full connection layer is removed and the reverse convolution layer is added to solve the problem that the size of the input candidate text box is not uniform.Then,the residual module is introduced to solve the problem of text detection in large-scale images;Finally,a loss function combining Dice coefficient loss function and balanced weight strategy is designed to make the trained network adapt to text detection of different scale images and meet the text detection requirements of relay protection equipment.In the text recognition part,a text recognition algorithm based on CNN and BLSTM(Bidirectional Long Short Memory Network)is designed.The results of the text detection part are input into the CNN based spatial feature extraction network,and the same feature extraction network structure is selected as the detection part.The extracted convolution features are converted into feature vectors,and then input into the BLSTM based semantic feature extraction network,After fully extracting the semantic features about the context,the output is sent to the output translation layer based on CTC model for translation,and the final text recognition result is obtained.