Li Guang, Liu Peng, Cao Xiaofan, Liu Chunguang
School of Control and Computer Engineering, North China Electric Power University, Beijing 102206, China.
Yangzhong Intelligent Electric Research Center, North China Electric Power University, Yangzhong 212211, China.
Sensors (Basel). 2023 Jun 14;23(12):5579. doi: 10.3390/s23125579.
Recently, hybrid Convolution-Transformer architectures have become popular due to their ability to capture both local and global image features and the advantage of lower computational cost over pure Transformer models. However, directly embedding a Transformer can result in the loss of convolution-based features, particularly fine-grained features. Therefore, using these architectures as the backbone of a re-identification task is not an effective approach. To address this challenge, we propose a feature fusion gate unit that dynamically adjusts the ratio of local and global features. The feature fusion gate unit fuses the convolution and self-attentive branches of the network with dynamic parameters based on the input information. This unit can be integrated into different layers or multiple residual blocks, which will have varying effects on the accuracy of the model. Using feature fusion gate units, we propose a simple and portable model called the dynamic weighting network or DWNet, which supports two backbones, ResNet and OSNet, called DWNet-R and DWNet-O, respectively. DWNet significantly improves re-identification performance over the original baseline, while maintaining reasonable computational consumption and number of parameters. Finally, our DWNet-R achieves an mAP of 87.53%, 79.18%, 50.03%, on the Market1501, DukeMTMC-reID, and MSMT17 datasets. Our DWNet-O achieves an mAP of 86.83%, 78.68%, 55.66%, on the Market1501, DukeMTMC-reID, and MSMT17 datasets.
最近，混合卷积-Transformer架构因其能够捕捉局部和全局图像特征以及相对于纯Transformer模型具有更低计算成本的优势而变得流行。然而，直接嵌入Transformer可能会导致基于卷积的特征丢失，特别是细粒度特征。因此，将这些架构用作重新识别任务的主干并不是一种有效的方法。为了应对这一挑战，我们提出了一种特征融合门单元，它可以动态调整局部和全局特征的比例。特征融合门单元根据输入信息，将网络的卷积分支和自注意力分支与动态参数进行融合。该单元可以集成到不同的层或多个残差块中，这将对模型的准确性产生不同的影响。使用特征融合门单元，我们提出了一种简单且可移植的模型，称为动态加权网络或DWNet，它支持两种主干，即ResNet和OSNet，分别称为DWNet-R和DWNet-O。DWNet在显著提高重新识别性能的同时，保持了合理的计算消耗和参数数量。最后，我们的DWNet-R在Market1501、DukeMTMC-reID和MSMT17数据集上分别实现了87.53%、79.18%、50.03%的平均精度均值（mAP）。我们的DWNet-O在Market1501、DukeMTMC-reID和MSMT17数据集上分别实现了86.83%、78.68%、55.66%的平均精度均值（mAP）。
