一种基于改进SegNet模型和融合注意力机制的苗族服饰图案分割算法  

A Miao Costume Pattern Segmentation Algorithm Based on Improved SegNet Model and Fusion Attention Mechanism

作　　者：谭永前[1] 曾凡菊[1] TAN Yong-qian;ZENG Fan-ju(Kaili University,Kaili,Guizhou 556011)

机构地区：[1]凯里学院,贵州凯里556011

出　　处：《怀化学院学报》2023年第5期66-71,共6页Journal of Huaihua University

基　　金：贵州省2022年度基础研究计划(自然科学)项目“基于卷积神经网络(CNN)和语义分割的苗族刺绣风格化迁移与保护研究”(黔科合基础-ZK[2022]一般526)。

摘　　要：针对目前语义分割模型需要消耗大量计算资源,对苗族服饰图案分割精度不高等问题,提出一种在改进的Seg Net模型中嵌入注意力机制模块的算法模型。改进后的Seg Net模型减少了卷积层数,在一定程度上减少了内存资源的消耗;为了提高对服饰图案细小纹理的分割精度,减少细小纹理细节的丢失,使用二分类交叉熵损失函数作为损失函数,提高了服饰图案细小纹理的分割精度;改进的Seg Net模型中嵌入轻量级的注意力模块,使得模型能够关注图像中更多的细节特征,更好地将感兴趣的特征从局部水平关联到全局水平,加强对目标特征信息的提取,减少损失率,提高模型的分割性能。实验结果表明,改进后的算法模型相对于传统相关算法,对苗族服饰图案的分割精度有了一定的提高,分割结果的视觉效果和定量指标上都要优于传统相关算法。This paper addresses issues such as the high computational resource consumption and low segmentation accuracy of existing semantic segmentation models for Miao ethnic clothing patterns.To overcome these challenges,an algorithm model is proposed by embedding an attention mechanism module into an improved SegNet model.The improved SegNet model reduces the number of convolutional layers,thereby reducing memory resource consumption to some extent.To enhance the segmentation accuracy of fine textures in clothing patterns and minimize the loss of intricate texture details,a binary cross-entropy loss function is employed.This loss function improves the segmentation accuracy of fine textures in clothing patterns.The improved SegNet model incorporates a lightweight attention module,allowing the model to focus on more detailed features within the image.This facilitates better integration of interested features from local to global levels,enhancing the extraction of target feature information,reducing loss,and improving the segmentation performance of the model.Experimental results demonstrate that the proposed algorithm model shows improved segmentation accuracy for Miao ethnic clothing patterns compared to traditional correlation-based algorithms.Both the visual effects and quantitative metrics of the segmentation results outperform those of traditional correlation-based algorithms.

关 键 词：图案分割 苗族服饰 注意力机制 SegNet模型 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]