改进DeepLabv3+的服饰图像分割算法

An Improved Garment Image Segmentation Algorithm based on Deeplabv3+

近年来,服装行业快速发展。少数民族服饰具有其独特的地域特征、审美价值以及使用价值,对其服饰的图像分割显得尤为重要。目前大多数传统图像分割基于深度卷积神经网络。其具有平移不变性、对位置不敏感、降采样过程会损失空间信息、不便于分割出结果等特点。为了更高效率分割少数名族服装图像,本文提出一种改进Deeplabv3+算法的少数民族服饰识别模型分割图片方法。本模型首先采用图片拼接数据增强方法、标签平滑、损失函数和交叉熵损失函数算法联合使用、辅佐分支结构、余弦退火函数等将少数民族服饰识别分割出来,然后采用了大量的少数民族服饰图片数据并对其划分为训练集和测试集进行实验。实验证明,该模型能够更有效地提高少数民族服饰识别的准确性和效率。

In recent years,the clothing industry has developed rapidly,especially the minority clothing has itsunique regional characteristics,aesthetic value and use value,so the image segmentation of its clothing is particularly important.At present,most traditional image segmentation is based on deep convolution neural network,which has the characteristics of translation invariance,insensitive to position,loss of spatial information during downsampling,and difficulty in segmentation results.In order to segment ethnic minority clothing images more efficiently,this paper proposes a method of image segmentation based on ethnic minority clothing recognition model with improved deeplabv3+algorithm The combination of dice loss and cross entropy loss,auxiliary branch structure,cosine annealing function,etc.are used to separate the ethnic costumes.Then a large number of ethnic costumes image data are used and divided into training sets and test sets for experiments.The experiments show that the model can more effectively improve the accuracy and efficiency of ethnic costumes recognition.