Research On Garment Image Semantic Segmentation Method Based On Deep Learning

As the carrier of ethnic cultural characteristics,ethnic minority costumes play an important role in traditional ethnic culture.The study of ethnic minority costume patterns is of great significance to the protection and inheritance of traditional ethnic culture.Because ethnic minority clothing image background complexity is low,with national costume jewellery is relatively similar,and so on,using the existing network segmentation technology prone to fitting,similar object segmentation task difficult technical problems,such as increased this text set about from the deep learning method,using convolution and multilayer perceptron neural network,Mixing attention mechanism in the network,and select the appropriate technology,such as loss function is proposed based on a deep study of ethnic dress image semantic segmentation method,this method improved technology,its network model is studied to solve the existing national costumes related technical problems that exist in the image segmentation method,improve the segmentation performance of network model,the main work done is as follows:(1)Design an Efficient-Deeplabv3+ network for Miao apparel recognition based on DeepLabv3+ network.In the network training,Mosaic data enhancement was used to increase the background complexity of training images to extract more feature information.Label smoothing is used to avoid network over-fitting due to over-reliance on training sample labels;The auxiliary branch structure is introduced to make full use of the information retained by the intermediate feature layer and calculate the loss value together with the main branch loss function.The joint loss function is used to calculate the loss value to prevent gradient explosion,and the multistage decay cosine annealing algorithm is used to find the optimal learning rate for the current number of iterations.The experimental results show that the average crossover ratio and the average pixel accuracy of categories of the Efficient-Deeplabv3+ network are 83.32% and 92.54%,respectively,in the Miao clothing data set.(2)The TGMLP-Efficient-DeepLabv3+ network is designed to improve the Efficient-DeepLabv3+ network.SE attentional mechanism is introduced in the network,and an efficient TGMLP structure is designed.Firstly,a TMLP module is designed.TMLP consists of three branches encoding along channel,width and high dimension respectively,which enables TGMLP not only to capture long-distance dependence of features along spatial dimensions but also to maintain accurate position information along the three-dimensional direction.Secondly,the Local and Global Perceptron modules are designed.The Global Perceptron module divides the feature map into different regions and conducts correlation modelling for each region to establish the Global dependency relationship between different regions.The Local Perceptron module utilizes the powerful Local feature extraction capability of convolution to process feature maps at multiple scales and explore the context information connections within structures.Finally,a gating mechanism is introduced in TGMLP to make it easier for the model to learn the position bias of feature maps in different scale data sets.The Focal Loss function was used to adjust sample weight and improve the segmentation performance of the network.Experimental results show that the TGMLP-Efficient-DeepLabv3+ network can obtain global context information,establish long-term dependence between features,segment objects more completely,and further improve network segmentation performance.The average crossed-over ratio and the average pixel accuracy of categories reached 84.96% and 93.7%,respectively.