面向可视着装分析的智能特征解析与潜在属性表征方法

This research emphasizes on the challenging theoretical problems in visual clothing analysis and their new practical applications. Considering the parsing, description and representation issues in the analysis of the clothing feature and style, we devote our efforts to study the intelligent feature parsing and latent attribute representation in visual clothing analysis. This research is highlighted by that it is to establish the novel scientific theories and solutions, based on the practical clothing circumstances and demands, with the potential to be applied in practices finally. The technical key points of the research are as follows. 1) Establish the conditional feature coupling scheme for multi-persons clothing parsing network. It could enhance the semantic recognition and the regional extraction in the complex solo and multi-persons scenes. 2) Design the fine-grained hierarchical description models related to the visual features and attributes. These models could raise the ability of the multi-scale feature description for the clothing styles by edge-structural, regional and attribute-based cues. 3) Implement the latent attribute representation framework for the clothing style prediction. It could boost the accuracy of the recognition for the fashion elements in clothes and satisfy the demands of new clothes applications. This research will develop a prototype system of visual clothing analysis to prove the proposed intelligent feature parsing and latent attribute representation models. It is expected to break some bottlenecks in the clothing analysis and provide a series of theoretical and technical supports for the new clothing applications as well.

项目针对可视着装分析理论及其新型服装应用所面临的挑战，围绕服装特征与风格分析过程中的解析、描述和表征问题，研究面向可视着装分析的智能特征解析与潜在属性表征方法。本项目研究特点是立足于现实的服装应用场景和需求，形成科学的理论与解决方案，再应用于实践之中。技术方法上重点研究：1）建立多种条件特征耦合策略下的多人分组着装解析框架，增强对复杂着装场景中的单人和多人着装语义分析和区域内容提取能力；2）设计可视特征与属性关联的精细分级描述模型，从边缘结构、区域部件和属性分级角度，强化对服装款式和着装风格的多层次特征刻画能力；3）实现针对着装风格预测的潜在属性表征框架，提升对服装中蕴含的时尚元素和风格要素的识别准确度，满足新兴服装应用的需要。该项目将通过实现面向可视着装分析过程的原型系统，验证上述提出的智能特征解析与潜在属性表征方法技术，并有望突破服装分析中的核心难题，为新型服装应用提供理论和技术支持。

项目根据服装时尚智能化的新需求，重点围绕着装特征及其风格元素中的解析、描述和表征问题，研究面向可视着装分析的智能特征解析与潜在属性表征方法。本项目的主要研究成果包括：（1）利用姿体部位信息增强的思想，结合多视图特征堆叠网络模式，提出具有局部和全局关联的部位表征增强的着装人体解析方法；（2）针对着装场景数据不平衡问题，提出了抑制数据标签冗余的操作方式，以及具有重平衡特性的人体解析模型；（3）面对服装属性表征及其应用需求，提出了基于属性感知异构图网络的服饰兼容度预测模型，并构造了深度多模态融合的服装风格检索方法；（4）为了满足对服装特征进行描述的需要，提出了基于十字交叉注意力模块的服装关键点检测方法，并形成具有结构一致性的虚拟试穿网络。该项目按计划如期完成，并通过实现面向可视着装分析过程的原型系统，验证所提出的智能特征解析与潜在属性表征方法技术，从而可为新型服装时尚应用奠定理论和技术基础。