融合视觉风格和标签约束的少数民族服装图像解析

Objective: The styles and structures of ethnic minority costumes are complex, and their visual styles vary. Due to factors such as the lack of semantic labels for ethnic costumes, complicated local features, and mutual interference among semantic labels, the accuracy and precision of image analysis of ethnic minority costumes are relatively low. Therefore, this paper proposes an image analysis method for ethnic minority costumes that combines visual style and label constraints. Methods: Firstly, based on the clothing image data set of 55 ethnic minorities constructed in this paper, the general semantic labels and ethnic semantic labels of ethnic minority costumes are customized according to the basic style structure, dressing area, accessories, and different visual styles. At the same time, 4 groups of annotation pairs are set, with a total of 8 annotation points. Then, combined with the custom semantic labels and training images with annotation pairs, the visual style is added to the deep fully convolutional neural network SegNet to fuse local and global features, and attribute prediction, style prediction, and triplet loss functions are introduced to preliminarily analyze the input images to be analyzed. Finally, the initially analyzed results are further optimized through the constructed label constraint network to avoid label interference, and the optimized final analysis results are obtained. Results: Verification was carried out on the constructed ethnic minority clothing image data set. The experimental results show that the annotation pairs effectively improve the detection accuracy of local features, the constructed visual style network can effectively fuse the global and local features of ethnic minority costumes, and the label constraint network solves the problem of mutual interference among labels. After combining the visual style network and the label constraint network, the average precision of ethnic minority costume analysis can be significantly improved, and the pixel accuracy reaches 90.54%. Conclusion: The image analysis method for ethnic minority costumes that combines visual style and label constraints proposed in this paper can improve the accuracy and precision of image analysis of ethnic minority costumes and has good significance for inheriting the motherland's culture and protecting intangible cultural heritage.

目的少数民族服装款式结构复杂,视觉风格各异。由于缺少民族服装语义标签、局部特征繁杂以及语义标签之间存在相互干扰等因素导致少数民族服装图像解析准确率和精度较低。因此,本文提出了一种融合视觉风格和标签约束的少数民族服装图像解析方法。方法首先基于本文构建的包含55个少数民族的服装图像数据集,按照基本款式结构、着装区域、配饰和不同视觉风格自定义少数民族服装的通用语义标签和民族语义标签,同时设置4组标注对,共8个标注点;然后,结合自定义语义标签和带有标注对的训练图像,在深度完全卷积神经网络SegNet中加入视觉风格以融合局部特征和全局特征,并引入属性预测、风格预测和三元组损失函数对输入的待解析图像进行初步解析;最后,通过构建的标签约束网络进一步优化初步解析结果,避免标签相互干扰,得到优化后的最终解析结果。结果在构建的少数民族服装图像数据集上进行验证,实验结果表明,标注对有效提升了局部特征的检测准确率,构建的视觉风格网络能够有效融合少数民族服装的全局特征和局部特征,标签约束网络解决了标签之间相互干扰的问题,在结合视觉风格网络和标签约束网络后,能够明显提升少数民族服装解析的平均精度,像素准确度达到了90. 54% 。结论本文提出的融合视觉风格和标签约束的少数民族服装图像解析方法,能够提高少数民族服装图像解析的准确率和精度,对传承祖国文化、保护非物质文化遗产具有很好的意义。