Research On Substation Scene Text Detection Based On Deep Learning

The detection of natural scene texts has always been a key issue in the field of OCR.The text detection and recognition work for actual scenes always requires specific solutions to achieve optimal results.This paper has conducted in-depth research on the text detection problem of substation scenarios.The text to be detected in the substation scene is mainly concentrated on the equipment nameplate.The text on the equipment nameplate contains the equipment name and equipment core parameter information.Compared with other natural scene text detection data,the text detection task of substation scene image faces the complex background,serious noise,and mixed Chinese and English problems.For these difficulties,the text starts from the network structure and improves the feature map resolution and enlarges the feature.The method of receptive field improves the text detection effect of the substation scene.The main work of this thesis includes: 1)According to the characteristics of large aspect ratio of text box,according to the characteristics of cavity convolution,two kinds of residual structure based on cavity convolution are designed,and a convolutional neural network with high resolution and large receptive field is proposed.The training of multiple receptive fields is realized,and the text detection performance of the network is improved.2)For the problem that the long text box contains small text box in the detection result,the non-maximum value suppression algorithm is improved.The algorithm introduces a new judgment index-the inclusion rate,according to the ratio of the area of the two text boxes.To judge the situation that the text boxes are included with each other,the improved non-maximum suppression algorithm obtains a more accurate position of the detection frame and improves the detection effect.3)According to the effect of text recognition,this paper further post-processes the obtained text box,proposes the text box merging strategy and text box stretching strategy,improves the recognition effect on the boundary text,and effectively improves the recognition of the text line.rate.The thesis is tested on the substation scene dataset,and the performance of the network structure and the improved non-maximum suppression algorithm is explored.The experimental results show that the expansion of the receptive field can effectively improve the effect of text detection.At the same time,the improved non-maximum suppression algorithm yields a more precise text box.