Precision Identification Method For Cards Captured From Natural Scenes

Many enterprise platforms adopt the model of "Mobile & Internet" to collect cards uploaded by users and automatically identify them.Some cards identification platforms have insufficient antiinterference ability to deal with factors such as illumination,perspective,complex background,watermark coverage,special printing,etc.,which strictly limit the quality of cards acquisition and brings inconvenience to users.Aiming at the identification of cards captured from natural scenes,this paper proposes three precise optimization methods for cards locating based on SIFT feature points matching,visible watermark removal based on Conditional Generative Adversarial Networks,and text recognition enhancement.Through experimental design,it is verified that the proposed method can effectively improve the efficiency and accuracy of content recognition for credential images with non-uniform illumination and complex background interference.The main work of this paper is as follows:First,a cards locating method based on SIFT feature points matching is proposed.Aiming at the problem of cards locating based on the invariance of the locating performance of the feature points of the custom template map,two locating point selection algorithms based on SIFT feature points detection,FLANN feature points matching and PROSAC error matching blanking are given:speed priority and accuracy priority.The real-time observation program of cards locating verifies the robustness of the method in complex environment changes.Secondly,using the pix2 pix model based on Conditional Generative Adversarial Networks,by proposing a large number of expansion methods for the paired training set,the goodness of fit of the training of the watermarking model is effectively improved.The content covered by the watermark is fully restored,and the anti-interference ability of the card text recognition is also enhanced.Finally,a method of recognition and enhancement of text in cards is proposed.On the one hand,through the field locating and character segmentation based on the "projection method",the clarification of printed text and the accelerated recognition of field merging,detailed optimization has been made at the level of field fine-grained purification and process recognition.On the other hand,in the generation of Tesseract-OCR custom language package,a sample generation method and automatic character correction algorithm are proposed based on the "projection method",which improves the success rate of Tesseract sample character detection and correction efficiency.Aiming at the identification of cards captured from natural scenes,this paper proposes a performance improvement plan from the above three aspects,which can provide application references for system development and implementation.