基于注意力机制与残差网络的掌纹脊线距离估计

Palmprint Ridge Distance Estimation Based on Attention Mechanism and Residual Network

掌纹脊线距离是掌纹脊线的重要纹理属性之一,在掌纹识别算法与图像质量评估方法中常作为一个重要参数被进行参考。目前常用的掌纹脊线距离估计方法为频域方法,但已有的频域方法对掌纹图像质量与完整性要求高,导致适用范围受限,且测量精度仍需进一步提高。针对上述问题,提出一种基于注意力机制与残差网络的掌纹脊线距离估计方法,将手工标注获取的脊线距离值作为输入掌纹图像块的类别进行识别。该方法以VGG16网络为基础,将CBAM注意力模块有机引入,增强网络对脊线结构信息的关注;并对损失函数进行改进,添加残差机制,来避免梯度消失问题。所用数据集由手工标注完成,针对数据集中存在的样本分布不均衡问题,采用数据增强的方法削弱类别分布间的不平衡,并设计样本不均衡损失函数对模型进行优化。实验结果表明,该方法在该数据集上相比频域方法在识别准确率上提升了14.2%,且相比部分经典深度学习方法在准确率上至少提升了5.1%。

Palmprint ridge distance is one of the important texture attributes of palmprint ridge.It is often referred as an important parameter in palmprint recognition algorithm and image quality evaluation method.At present,the commonly used method to estimate palmprint ridge distance is frequency domain method,but the existing frequency domain method has high requirements on the quality and integrity of palmprint image,resulting in limited application scope and further improvement of measurement accuracy.To solve these problems,a palmprint ridge distance estimation method based on attention mechanism and residual network is proposed.The ridge distance value obtained by manual annotation is recognized as the category of the input palmprint image block.Based on VGG16 network,CBAM attention module is introduced to enhance the network's attention to ridge structure information.The loss function is improved and the residual mechanism is added to avoid the gradient disappearance.The data set is manually annotated.For the unbalanced distribution of samples in the data set,we employ the method of data enhancement to weaken the imbalance between the class distributions,and design the sample unbalanced loss function to optimize the model.Experimental results show that the proposed method has a 14.2%improvement in recognition accuracy compared with the frequency domain methods on this dataset,and at least a 5.1%improvement in recognition accuracy compared with some classical deep learning methods.