Minority Costume Recognition Based On Multi-scale Attention Mechanism

Ethnic costumes are different from ordinary costumes that people are familiar with everyday.Ethnic costumes have their own unique styles and patterns,which can most directly and clearly express the culture and customs of ethnic minorities.Therefore,in all aspects of ethnic minority culture and customs,etc.All have great research value.Based on the deep learning method,the thesis conducts research on ethnic clothing.The color and detailed feature information contained in ethnic clothing images is relatively rich.Moreover,the clothing of the same ethnic group is very different in appearance and color.It is impossible to use a single feature for ethnic clothing.Describe and identify.At the same time,clothing pictures are affected by lighting,background and clothing deformation.Traditional image extraction and classification methods have general effects,combined with multi-scale feature extraction and attention mechanism methods to achieve the classification of ethnic minority clothing.The main work content is as follows:First,construct an image collection of ethnic costumes,a total of 9 types of images,a total of 1,607 images.Including the Korean,Bai,Dai,Manchu,Mongolian,Miao,Tibetan,Uygur,and Zhuang ethnic minority characteristic images in 9 categories.Effectively solve the problem of the lack of data sets for the identification of ethnic minority costumes.Second,in response to the current problem of the richer details of ethnic clothing recognition,the paper introduces multi-scale dense connection units into the Dense Net network,constructs a multi-scale ethnic clothing recognition network based on Dense Net,and realizes the task of multi-ethnic clothing recognition.Compared with previous algorithms,the multi-scale densely connected unit can extract the detailed features of ethnic clothing from multiple scales and dimensions,which greatly improves the recognition rate of clothing.Thirdly,in view of the problem that ethnic costumes are affected by background light,the paper introduces the concept of attention,and realizes a method of ethnic costume recognition based on attention mechanism.Through the acquisition of local and global attention mechanisms,the network is more inclined to focus on information that can help judgment in the image,and ignore irrelevant information.And add the visualization of the attention mechanism to prove the effectiveness of the attention mechanism.The method proposed in the paper achieves a 95.18% accuracy rate in the classification of ethnic minority clothing,and proposes a local and global attention mechanism,which enables the network to focus on the part more relevant to recognition,ignoring the background and illumination.Interference.