NexusNet：一种面向面部颜色识别的双路径层次化融合混合网络

NexusNet: A Dual-Path Hierarchical Fusion Hybrid Network for Facial Color Recognition

摘要: 针对当前卷积神经网络(CNN)在建模长距离依赖上的局限，以及视觉Transformer因自注意力机制导致的参数量庞大问题，本文提出一种双路径混合模型——NexusNet。该模型通过深度融合CNN的局部表示分支与Transformer的全局建模分支，实现了局部细节特征与全局语义信息的协同编码，在显著提升特征表征能力的同时保持了精简的参数量。在CNN分支中，我们引入了融合动态权重分配与上下文增强机制的新型模块，以增强对判别性局部结构的捕捉能力；在Transformer分支中，采用分层建模与线性复杂度设计，大幅降低了长距离依赖建模的资源开销。此外，设计了一种自适应多层次特征融合模块，通过通道与空间注意力引导的多尺度特征整合，实现跨架构信息的高效聚合与参数优化。在两个面部颜色识别数据集上的实验表明，NexusNet在保持模型轻量化的前提下，分类准确率分别达到88.99%和79.25%，并在多项评价指标上优于现有主流方法，验证了其在局部‑全局特征融合与模型轻量化方面的有效性与泛化能力。

Abstract: To address the limitations of Convolutional Neural Networks (CNNs) in modeling long-range dependencies and the high parameter complexity of Vision Transformers, this paper proposes a dual-path hybrid model, NexusNet. The model integrates a CNN-based pathway for local feature extraction with a Transformer-based pathway for global context modeling, enabling effective fusion of fine-grained details and semantic information while maintaining model compactness. In the CNN pathway, we introduce a novel module that combines dynamic weight allocation with a context enhancement mechanism to improve discriminative local feature capture. The Transformer pathway employs a hierarchical structure with linear complexity to efficiently model long-range dependencies. Furthermore, we design an adaptive multi-level feature fusion module that leverages both channel and spatial attention to guide the integration of multi-scale features from both architectures, promoting efficient information aggregation. Experimental results on two facial color recognition datasets demonstrate that NexusNet achieves classification accuracies of 88.99% and 79.25%, respectively, and outperforms existing methods across multiple metrics. This validates the model’s strong performance and generalization ability in joint local-global representation learning and efficient model design.