| 摘 要: 针对全成形针织图案逆向工程中存在的特征纠缠、拓扑信息丢失和极端的长尾分布难题，提出一种基于双分支特征解耦与拓扑约束的针织指令生成网络（Instruction-Aware Dual-Branch Network, IADBN）。首先，构建物理属性解耦的双分支特征提取架构，利用空洞卷积与标准卷积分别提取全局结构与局部细节特征，解决微弱纹理被背景淹没的问题；其次，引入拓扑感知坐标注意力模块（TA-CA），通过分解特征聚合显式保留针床行列位置信息，强化网格拓扑约束；最后，设计融合平衡Softmax与逻辑兼容性约束的混合损失函数，修正长尾分布偏差并降低逻辑错误率。实验结果表明，该方法在MIT标准数据集上的整体准确率达到95.24%，极度稀有指令识别准确率较现有最优方法提升10%以上，此外，结合多维度数据增强策略，模型在光照偏移与高斯噪声等复杂工况下的性能衰减均严格控制在1%以内。生成的指令图具备更高的工程实用价值与真实场景泛化能力。 |
| 关键词: 深度学习 针织指令生成 双分支网络 坐标注意力 长尾分布 |
|
中图分类号:
文献标识码:
|
| 基金项目: 浙江省高层次人才特殊支持计划（2023R5212）；国家重点研发计划项目：纱线生产关键工序质量在线检测传感器及系统应用（SQ2023YFB3200093） |
|
| Knitting Instruction Generation Method Based on Deep Learning Fusing Dual-Branch Feature Decoupling and Topological Constraints |
|
YANG Fan1, PENG Laihu1, QI Yubao1, WEI Yin2, SONG Kejia2
|
1.Zhejiang Sci-Tech University;2.Zhejiang Rifa Textile Machinery Company Limited
|
| Abstract: To address the critical challenges in the reverse engineering of whole-garment knitting patterns—specifically feature entanglement, loss of topological information, and extreme class imbalance—this paper proposes an Instruction-Aware Dual-Branch Network (IADBN) based on dual-branch feature decoupling and topological constraints. Firstly, a physically decoupled dual-branch feature extraction architecture is constructed, utilizing dilated convolutions and standard convolutions to extract global structural features and local detail features, respectively, thereby resolving the issue of weak textures being submerged by the background. Secondly, a Topology-Aware Coordinate Attention (TA-CA) module is introduced to explicitly preserve the absolute positional information of the needle bed by decomposing feature aggregation, thus reinforcing grid topological constraints. Finally, a hybrid loss function integrating Balanced Softmax and logical compatibility constraints is designed to correct biases derived from long-tailed distributions and reduce logical error rates. Experimental results on the standard MIT dataset demonstrate that the proposed method achieves an overall accuracy of 95.24%. Notably, the recognition accuracy for extremely rare instructions is improved by over 10% compared to existing state-of-the-art methods. Furthermore, combined with multi-dimensional data augmentation strategies, the performance degradation of the model under complex conditions such as brightness shift and Gaussian noise is strictly controlled within 1%. This indicates that the generated instruction maps possess significantly higher engineering practical value and generalization ability in real-world scenarios. |
| Keywords: Deep learning Knitting instruction generation Dual-branch network Coordinate attention Long-tailed distribution |
