Seal Detection For Document Images Based On Data Synthesis And Multi-mask Loss Function

Seal is a very symbolic object used for signing or authentication.With the popularization of text detection technology,there is an urgent need for the detection of seals in document images.Traditional seal detection methods mostly use gradient segmentation to analyze the outline of the seal,the extraction of color feature components,and the feature matching based on the seal template to achieve the detection of the seal.However,when the traditional seal detection methods face the complex background with more noises,the color of the seal and the background are close,and one image with multiple seals,etc.Their detection effect will be greatly reduced.This paper proposes a deep learning model of multi-mask loss function for seal detection.Five masks are designed for the key information points in the seal to semantically segment the seal.It can not only locate and detect the seal,but also the key words such as the text of the seal.Aiming at the scarcity of data sets,this paper synthesizes sufficient seal sample data through data synthesis technology,and expands and enhances the data to ensure the diversity of data and the robustness of the model by simulating many situations that may occur in natural scenes on the basis of synthetic data.Before detection,digital image processing technology is used to preprocess the image to further improve the detection accuracy.Compared with the traditional method,the method proposed in this paper can better deal with the real situation such as seals without color characteristics and one picture with multiple seals,and has higher accuracy and better experimental results for the detection of seals and their texts.The main contributions of this article are as follows:(1)Apply data synthesis technology to seal dete ction.Due to the scarcity of seal data,there is currently no official data set.This paper synthesizes the data set of seal detection by data synthesis,and adds various interference factors that may appear in real life to the synthesized data to make th e data more authentic and diverse.The test set not only contains synthetic data with a variety of possible morphological transformations and noises,but also contains hand-labeled real data,making the model more close to the real natural scene,thereby solving the bottleneck problem of data scarcity.(2)Preprocess the document image with the related technology of image processing.Properly extract and denoise the seal area and strengthen the main part of the seal to a certain extent to improve detection accuracy and efficiency.(3)Propose and implement a deep learning model of multi-mask loss function for seal detection.The model designs corresponding masks for five key points: the position of the seal,the type of the seal,the position of the text,the position of the seal logo,and the apex of the seal.The seal in the document image is semantically segmented through each mask,and the document is finally completed seal detection in the image.