Abstract: In order to solve the problem of feature information loss in Miao apparel image segmentation based on the deep learning method, an Efficient-DeepLabv3+ network for Miao apparel recognition was designed. Firstly, Mosaic data enhancement increased the background complexity of images during training so that the network can extract more image feature information without the additional computational overhead. Secondly, label smoothing was used to reduce actual label training loss weight and reduced the adverse effects of over-fitting on the segmentation effect. Thirdly, the auxiliary branch structure was introduced so that the loss function can calculate the loss value of all network layers. In order to prevent gradient explosion and make network training more stable, a joint loss function was proposed to calculate the loss value. Finally, a multistage attenuation cosine annealing algorithm was proposed to find the global optimal learning rate and speed up the convergence of network training. The experimental results show that Mean Intersection over Union (MIoU) and category average Pixel Accuracy (MPA) reach 84.96% and 93.7%, respectively, on the Miao clothing data set. On the PASCAL VOC2012 data set, the segmentation effect of Efficient-DeepLabv3+ network is better than other networks.