Research And Application Of Salient Mongolian Pattern Detection Method Based On Deep Learning

As a traditional art form with distinctive regional and ethnic characteristics,Mongolian motifs are an important part of Mongolian culture.It manifests the spiritual character of the Mongolian people and carries on the aesthetic values of the Mongolian people.Most of the traditional Mongolian motifs are mainly combination motif patterns,and the conspicuous areas in the combination motifs represent the artistic form and meaning of the motifs,which are the main parts extracted by the non-genetic inheritors of Mongolian motifs.However,there is a lack of intelligent processing tools,and the extraction of such patterns can only be done manually by non-hereditary bearers.Therefore,the extraction of significant Mongolian patterns suffers from long extraction cycles,low efficiency,and the extraction process is easily influenced by personal subjective factors,resulting in the lack of an accurate and reliable basis for determining the extraction results and making it difficult to produce and apply them on a large scale and in bulk.With the rapid development in the field of computer vision,deep learning-based saliency target detection models are widely used in many fields.Mongolian traditional patterns differ significantly in size,structure and colour distribution compared to images in existing datasets such as portraits and natural images,and there is a lack of datasets for salient Mongolian patterns,so existing methods cannot be directly applied to Mongolian pattern detection.The prediction results of salient Mongolian motifs require strict integrity and detail,and existing saliency target detection models are inadequate in focusing on multi-scale image contextual information and predicting salient graph details.An intelligent detection method for salient Mongolian motifs is a key step in the reproduction and exploitation of Mongolian motifs,which will help Mongolian motif resources to be used efficiently,enhance the competitiveness of related products in the market,and promote the heritage and development of Mongolian motifs.This thesis focuses on a deep learning-based method for detecting salient Mongolian motifs,designed and implemented a salient Mongolian pattern detection and processing system,with the following work:(1)Constructing a significant Mongolian pattern detection dataset.To address the problem of missing datasets,a dataset containing 1900 Mongolian patterns was constructed in four steps:collection,pre-determination,labeling and expansion,which fills the gap of significant Mongolian pattern detection datasets in the field of deep learning and lays the foundation for the subsequent in-depth study of Mongolian patterns.(2)A salient Mongols pattern detection algorithm named SFEMAU~2-Net is proposed.The network has added supplementary feature extraction modules and lightweight combination attention modules in the encoder stage of U~2-Net,improving the model‘s ability to supplement and fuse multi-level features.At the same time,it enhances the model‘s attention to effective spatial and channel information,thereby improving the integrity of the prediction of significant Mongolian patterns.Comparative experiments were conducted with 8 advanced methods such as Res2Net on 4 commonly used public datasets in the field,and the experimental results showed that SFEMAU~2-Net has a certain degree of universal effectiveness.At the same time,a comparative test was conducted with four advanced open source methods,such as Pool Net,on the self-made Mongols pattern dataset,and the three evaluation indicators all achieved the best results,fully proving the field optimality of SFEMAU~2-Net.In addition,our own ablation experiment has been added.According to the results of the ablation experiment,when introducing the new module in this article,the MAE index decreased by 0.0149 and max F compared to U~2-NetβThe indicators and Em indicators have increased by 0.014and 0.028 respectively compared to U~2-Net,fully proving the effectiveness of the proposed network structure components.(3)Designed and implemented a salient Mongolian motif detection and processing system.By embedding the SFEMAU~2-Net method and various image processing algorithms,the functions of online extraction of salient Mongolian motifs,image cropping and editing,and automatic generation of line drawings are realized.It provides a convenient and efficient tool for the high-quality reuse of digitised Mongolian motifs,simplifies the processing of Mongolian motifs,and contributes to the innovation and development of Mongolian traditional culture.