基于改进的DenseNet-BC对少数民族服饰的识别

Minority clothing recognition based on improved DenseNet-BC

随着信息技术的发展,数字技术越来越多地应用于民族文化数字化保护,民族服饰的数字化及分类问题也日益受关注。相比一般服饰,少数民族服饰具有更多的细节特征信息,对其进行分类识别具有很大挑战。选用卷积神经网络DenseNet-BC作为基础网络结构,设计并使用了多尺度密集连接单元,用不同大小的卷积提取不同尺度的特征信息,以提高网络的学习能力;此外,为进一步提高网络的鲁棒性,提出一种局部和全局注意力机制方法进行分类识别。实验结果表明,改进的DenseNet-BC模型对少数民族服饰的识别准确率达95.18%,较ResNet-18、ResNet-34和DenseNet模型的识别准确率分别提升了3.84%、2.27%和1.18%。改进的DenseNet-BC模型具有更好的特征提取能力,能够提取更多的细节特征信息,一定程度上解决了普通模型提取特征尺度单一、特征丰富度低的问题。

With the development of information technology,the digital protection of national culture has received more and more attention,the digitalization and classification of national costumes have also become a hot topic.Compared with general clothing,ethnic minority clothing holds more detailed feature information,which poses a big challenge to classify and identify them.This paper selects DenseNet-BC as the basic network structure,and uses multi-scale dense connection units to extract feature information at different scales through different convolution sizes hence improving the learning ability of the network.Furthermore,a local and global attention mechanism is proposed for classification and recognition to make the network robust.Experimental results show that the recognition accuracy of the improved DenseNet-BC model is 95.18%,which is respectively 3.84%,2.27%,and 1.18%higher than the recognition accuracy of ResNet-18,ResNet-34 and DenseNet models.The improved DenseNet-BC model has also better feature extraction capabilities,and can extract more detailed feature information.In brief,this network solves partially the problem of single feature scale and low feature richness of common models.