Clothing Pattern Structure Modeling and Reconstruction via Multi-Module Fusion Graph Neural Networks with Path Planning and Reinforcement Learning

Abstract

There are core difficulties in the intelligent recognition and generation application of clothing pattern structure, such as irregular geometric topology, weakened semantic structure, and unstable path planning. To solve such problems, an intelligent feature extraction and structure reconstruction path learning scheme that integrates graph neural networks is constructed. In the stage of structural diagram modeling, a clothing structure diagram is constructed based on the node edge surface configuration relationship. The graph convolutional network is used to embed the spatial adjacency relationship in multiple dimensions, supplemented by attention mechanism to enhance the response ability of key nodes and improve the stability of extracting local salient features. To better express the relationship between structural semantics and geometry, a multi-scale graph embedding strategy and structural context aggregation module are introduced to enable nodes to have stronger expressive power in both topological and semantic dimensions. In terms of reconstructing path generation, a graph autoencoder architecture is introduced to achieve controllable mapping of structure to path space, integrating geometric consistency constraints to enhance structural accuracy. The path decision-making process adopts a reinforcement learning model based on policy gradient, and optimizes the path guidance process through feedback mechanism. This experiment is based on the DeepFashion2 public dataset and our self built clothing structure graph data, with a total of 4826 samples and an average of 43 vertices. The results show that the accuracy index of our model reaches 91.3%+0.5, the Topology Score reaches 88.0%+0.6, and the F1 Structure Score reaches 88.4%+0.6, which is much higher than the basic method. The innovation of this study is mainly reflected in three aspects: proposing the use of graph convolution+attention to achieve multi task feature extraction; Introducing geometric constraints and policy networks to achieve reconstruction methods that maintain path consistency; The first application of GNN in the establishment of clothing style structure brings a new approach compared to traditional graph mapping.