基于SAM掩膜增强的服装解析方法

Clothing Parsing Method Based on SAM Mask Enhancement

服装解析任务旨在分割图像中服装实例的同时识别实例属性,是服装智能化研究中的主要任务之一。现有方法存在实例边界精度低,训练收敛速度慢的问题。得益于SAM大模型优秀的图像分割性能,设计了一种基于SAM掩膜增强的服装解析方法。首先提取输入服装图像的特征图、初始掩膜和边界框;然后,将边界框和特征图输入SAM解码出SAM掩膜,并与初始掩膜融合获得增强掩膜;最后,增强掩膜与特征图聚合得到实例查询的特征,并动态更新预测分支的权重,提高预测性能。与现有方法相比,该方法分割服装边界更加精确,模型训练收敛速度更快,在AP_(IoU)^(mask)评估指标上取得了最好的结果。

The clothing parsing task aims to segment clothing instances in an image while recognizing instance attributes,which is one of the main tasks in clothing intelligence research.Existing methods have the problems of low instance boundary accuracy and slow training convergence.Thanks to the excellent image segmentation performance of SAM large model,this paper designs a garment parsing method based on SAM mask enhancement.The method first extracts the feature map,initial mask and bounding box of the input clothing image.The bounding box and feature map are input into SAM to decode the SAM mask and fused with the initial mask to obtain the enhancement mask.Finally,the enhancement mask is aggregated with the feature map to obtain the features of the instance query and the weights of the prediction branches are dynamically updated to improve the AP_(IoU)^(mask) prediction performance.