The Tolerance Contradiction And Its Reconciliatory For Seal Difference Between The Presentation And Rationality

Since ancient times,seals have always been a representative of personal identity and collective reputation with legal effect,playing an important role in documents such as checks and contracts.Therefore,the recognition of seals is also of great importance.This thesis mainly studies the circular seal with a five-pointed star in the middle,provides solutions to two kinds of seal images with changing conditions,and introduces convolutional neural network into the seal recognition,proposes a part-to-holistic seal recognition idea,which provides theoretical guidance for the reconciliation of seal difference detection and seal rationality tolerance contradiction.The specific research content is as follows:(1)The basic methods of seal extraction and preprocessing are introduced,and two methods of seal extraction based on color space are compared,the method suitable for seal extraction in checks is obtained.Aiming at the seal images generated by using different stamping strength and angle,a seal image enhancement scheme is proposed.The image is processed by using histogram regulation and gray level equalization to rationalize the seal image and facilitate subsequent recognition.(2)Collect the seal image dataset.The collected data is composed of manually stamped images.In the process of stamping,various stamp images are obtained by using different stamping methods to make the data reliable.According to the characteristics of seal image includes rich features and dense texts,an idea of segmentation and recognition is used.According to the characteristics of seal border and center mark,the seal was initially recognized.Aiming at the feature of differences between text and background pixel values,a text segmentation technology based on the contour search method is proposed,which realizes complete segmentation of text in the seal.Since the extracted text image may contain the edge information of surrounding texts,the impurities are removed according to the angle of text in the seal.Finally,the text extraction without impurities is achieved.(3)Research the seal recognition technology based on deep learning.The collected seal image dataset is segmented into text images to generate a separate text dataset,and annotation information in a specific format is produced.The overall dataset is divided into three parts: training set,validation set and test set,which including 7500 samples and the same amount of labeled information.The Alex Net,VGG16 and Res Net50 three networks are trained and tested,and the training process of the network is optimized by transfer learning.Compare results of three optimized networks,select the best VGG16 network to test single text images,and the recognition accuracy can reach 96%.(4)The authenticity of the seal is judged by the result of text label matching,and the algorithm results of the key steps are displayed in the form of an interface on Qt platform.First of all,the segmented texts are sorted clockwise so that the final output text labels have a certain order.Secondly,seven different types of fake seals are identified,and the authenticity of the seal is judged according to the recognition results of the texts.Finally,the real seals images under different stamping conditions are tested,and the recognition accuracy before and after adjustment are compared.The adjusted recognition accuracy can reach 92.3%,and the detection rate is 91.4% for the data set composed of 35 false seal images.