基于自适应特征提取网络的复杂环境下人脸识别

中图分类号:TP391.41

摘要:

针对现有人脸识别算法在运动模糊、低光照等真实复杂环境下识别率低、鲁棒性较差，导致难以稳定应用在实际人脸识别任务的问题，提出一种基于自适应特征提取网络的复杂环境下人脸识别方法；该网络结合传统方法的特征提取技术和深度学习网络特征表示能力，实现了对不同复杂环境下人脸稳定识别；设计了一种自适应纹理特征提取算法，通过自动获取阈值来实现特征提取，提高网络计算效率；使用逆向传播算法改进深度信念网络，并引入共轭梯度算法解决网络的梯度消失问题，减少其收敛时间，提高算法鲁棒性；经实验验证，所提方法在标准LWF数据集和复杂环境CASIA、MS1M数据集中的准确率分别达到99.72%、89.54%及88.75%，参数量和网络计算量分别为2.84M及0.67G，均优于对比算法，能够满足复杂环境下人脸识别任务需求。

Abstract:

Aiming at the problem that the existing face recognition algorithms have low recognition rates and poor robustness in real and complex environments such as motion blur and low light, which makes it difficult to be stably applied to actual face recognition tasks, a face recognition method in complex environments based on adaptive feature extraction network is proposed. The network combines the feature extraction technology of traditional methods with the feature representation ability of deep learning network, and realizes the task of stable face recognition in different complex environments. An adaptive texture feature extraction algorithm is designed, which realizes feature extraction by automatically obtaining the threshold value and improves the network computing efficiency. The back propagation algorithm is used to improve the deep belief network, and the conjugate gradient algorithm is introduced to solve the gradient disappearance problem of the network, which reduces its convergence time and improves the algorithm's robustness. The experimental results show that the accuracy of the proposed method reaches 99.72%, 89.54% and 88.75% respectively on the standard LWF dataset and the complex environment CASIA and MS1M datasets. The number of parameters and network calculations are 2.84M and 0.67G respectively, which are superior to the comparison algorithm and can meet the needs of face recognition tasks in complex environments.