Research On Text Detection And Recognition Algorithm Of Fuzzy Inspection Image

With the vigorous development of machine vision technology,domestic inspection work is gradually in line with the field of machine vision.As an important organ performing the supervision function of national laws,it is not only necessary to transfer procuratorial documents within the organ,but also to accurately and efficiently transfer information and documents with Public Security Bureau,court and other political and legal organs.However,most of the relevant inspection materials are paper materials,so it is difficult to quickly obtain the effective information in the inspection image.Convert the document materials into digital information and store it on the inspection big data platform,so as to better assist the inspection work and promote the implementation of intelligent inspection work.At present,the more common text information extraction technology is Optical Character Recognition,but OCR technology has high requirements for the shooting environment,shooting angle and definition of the picture.If the picture is blurred or stained,the text recognition effect will be very poor.This paper focuses on the image features presented by different illumination and different blur types for inspection images with different blur degrees.Based on biological immunity,maximum stability threshold and convolution recurrent neural network,three fuzzy inspection text target detection and recognition algorithms are designed to complete the correct and rapid recognition of inspection text.The inspection text data set collected in this paper is used as the experimental material for simulation experiment analysis.The main research contents are as follows:(1)Aiming at the blur types such as slight blur,uneven illumination,and low color rendering,an adaptive immune factor algorithm is proposed.The algorithm is inspired by the immune process of biological cytology,combined with the artificial immune algorithm,and applies the immune process mechanism of the human body to the text target extraction process of fuzzy inspection images.Different immune factors are designed to extract text from inspection images with different fuzzy types.Experimental results show that the algorithm can improve the integrity and accuracy of text target extraction.(2)Aiming at the problem of deep fuzzy inspection image caused by ink stains of different colors covering the text area of inspection image,a text extraction algorithm based on local mapping maximum stable extremum is proposed.Firstly,the gray features in the local range are nonlinear mapped to expand the gray range;Then,the stain area is found by the joint location of the marking matrix and the maximum stable extremum region algorithm;Finally,the stained part is divided into sub images for text extraction alone.Through the correction and combination of sub-region texts,the final extraction result is obtained.The experimental results show that the algorithm has high accuracy and feasibility for the extraction of deep fuzzy inspection text.(3)In order to improve the accuracy and recognition time of fuzzy prosecution image recognition,a new inspection text recognition algorithm based on CRNN(Convolutional Recurrent Neural Network)model under the condition of small samples is proposed.In this paper,while ensuring the recognition accuracy,the timeliness of text recognition is improved;the deep bidirectional GRU(Gated Recurrent Unit)is used to replace the original deep bidirectional Long Short Term Memory.In this paper,we use the method of transfer learning to initialize the partial weight of CRNNBi LSTM in the large sample dataset and initialize the network to solve the problem of less training samples.Experiments show that the algorithm has a high recognition accuracy and effectiveness in deep fuzzy image detection tasks.In addition,this paper combines several proposed algorithms,designs and develops a text recognition software by using Py Side2 and QT Designer to realize real-time text detection.