计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (3): 175-183.DOI: 10.3778/j.issn.1002-8331.2109-0039

• 图形图像处理 • 上一篇    下一篇

基于跨批次存储预训练的素描人脸识别方法

邵玉颖,曹林,康峻,宋沛然,杜康宁,郭亚男   

  1. 1.北京信息科技大学 信息与通信工程学院,北京 100101
    2.北京信息科技大学 光电测量技术与仪器教育部重点实验室,北京 100101
    3.中国科学院 空天信息创新研究院,北京 100080
  • 出版日期:2023-02-01 发布日期:2023-02-01

Face Sketch Recognition Method Based on Cross-Batch Memory Pre-Training

SHAO Yuying, CAO Lin, KANG Jun, SONG Peiran, DU Kangning, GUO Yanan   

  1. 1.School of Information and Communication Engineering, Beijing Information Science and Technology University, Beijing 100101, China
    2.Key Laboratory of the Ministry of Education for Optoelectronic Measurement Technology and Instrument, Beijing Information Science and Technology University, Beijing 100101, China
    3.Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100080, China
  • Online:2023-02-01 Published:2023-02-01

摘要: 素描人脸识别技术在刑侦领域应用广泛,有助于缩小嫌疑人的搜寻范围。由于素描人脸样本数量不足,导致经典的深度学习模型无法达到理想的识别精度。针对此问题,提出一种基于跨批次预训练的素描人脸识别方法,通过在有限素描人脸数据集进行跨批次预训练的方式缓解训练样本稀缺问题,从而提高人脸识别模型的泛化能力。该方法通过跨批次存储机制缓解GPU存储限制扩大单批次预训练样本数量,从而获得更优的模型初始参数,并在其基础上根据三元组损失进一步优化模型,以提升网络性能。提出的方法在UoM-SGFS素描人脸数据集上的Rank-1识别精度为72.53%,在PRIP-VSGC数据集上Rank-10识别精度为62.47%。相比CDAN、DANN、SSD等方法识别率有显著提高。

关键词: 跨批次存储, 素描人脸识别, 预训练, 样本稀缺, 批次大小

Abstract: Face sketch recognition technology is widely used in the field of criminal investigation. It helps to narrow the scope of suspect search. Due to the insufficient number of sketched face samples, the classic deep learning model cannot achieve the ideal recognition accuracy. In response to this problem, this paper proposes a face sketch recognition method based on cross-batch pre-training. The problem of scarcity of training samples is alleviated by performing cross-batch pre-training on a limited sketch face data set. Thereby it improves the generalization ability of the face recognition model. This method relieves the GPU storage limitation through the cross-batch storage mechanism and expands the number of pre-training samples in a single batch, so as to obtain better initial model parameters. On this basis, the model is further optimized according to the triple loss to improve the network performance. The proposed method has a Rank-1 recognition accuracy of 72.53% on the UoM-SGFS sketch face dataset, and a Rank-10 recognition accuracy of 62.47% on the PRIP-VSGC dataset. Compared with CDAN, DANN, SSD and other methods, the recognition rate has been significantly improved.

Key words: cross-batch storage, face sketch recognition, pre-training, sample scarcity, batch size