SFEMAU2-Net: An Improved Saliency Mongolian Pattern Detection Algorithm for U2-Net

As a traditional art form with distinctive regional and national characteristics, Mongolian partterns are an important part of Mongolian culture. Most of the traditional Mongolian patterns are combination patterns, and the prominent parts of the combination patterns represent the art forms and meanings of the patterns. However, at present, the extraction of this kind of patterns can only be accomplished by the manual extraction of the intangible cultural heritage inheritors, so the extraction of the important Mongolian patterns has the problems of long extraction period, low efficiency, and the extraction process is easily affected by the individual subjective factors, and so on. To address this situation, this paper constructs a dataset for training and testing the important Mongolian pattern detection model, and proposes a salient Mongolian pattern detection algorithm SFEMAU2-Net (Supplementary feature extraction Attention U2-Net), which adds a feature Supplementary extraction module and lightweight combined attention module, which alleviates the problems of feature loss and unclear details of salient regions caused by multiple down-sampling in the feature extraction stage of the original U2-Net. The experimental results show that the proposed algorithm performs well on both the public dataset and the self-created Mongolian pattern dataset, which is very conducive to the research and promotion of Mongolian traditional culture.