基于特征掩膜的局部遮挡牛脸识别方法

Feature Mask-based Local Occlusion Cattle Face Recognition Method

摘要:

随着智慧牧业的高速发展，牛脸识别已成为牛场智能化养殖的关键，但现实应用场景中牛脸遮挡问题较为严重，影响识别系统的性能。为此，提出一种遮挡物分割辅助牛脸识别的全新双分支网络结构。首先设计一种改进的轻量级U-Net遮挡物分割模型，通过加入深度可分离卷积和多尺度混合池化模块，有效提高分割网络对遮挡物的分割性能。为更好地衰减遮挡物对牛脸识别性能的影响，引入一种多级掩膜生成单元。以不同层级的遮挡分割为输入，构建识别网络不同阶段所对应的掩膜，通过掩膜运算在特征提取的各阶段有效消除遮挡造成的损坏特征信息。最后在自制数据集上进行算法有效性和实时性验证，并与多种最新的典型识别算法进行对比。实验结果表明，本文算法在遮挡牛脸数据集上平均准确率达86.34%，识别速度为54f/s，且在不同程度遮挡的场景下，识别效果均优于FaceNet网络。

Abstract:

With the rapid development of intelligent animal husbandry, bovine face recognition has become the key to intelligent cattle breeding, but the problem of bovine face occlusion in practical application scenarios is more serious, which brings challenges to the performance of the recognition system. To solve this problem, a two-branch network structure based on occlude-assisted bovine face recognition was proposed. Firstly, an improved lightweight U-Net occlusion segmentation model was designed. By adding deep separable convolution and multi-scale mixing pool module, the occlusion segmentation performance of the segmentation network was effectively improved. Secondly, in order to better attenuate the influence of occlusions on bovine face recognition performance, a multilevel mask generation unit was introduced, and masks corresponding to different stages of the recognition network were constructed with different levels of occlusions as input. The damaged feature information caused by occlusions was effectively eliminated in each stage of feature extraction through mask operation. Finally, for the validity and real-time performance of the detection algorithm, the algorithm was verified on the selfmade data set, and compared with a variety of recent typical recognition algorithms. The experimental results showed that the proposed algorithm had an average accuracy of 86.34% on the blocked cow face data set, and the recognition speed was 54 f/s. Compared with the single-scale mask, the average accuracy of multistage mask was improved by 2.02 percentage points, and the recognition effect was better than that of the comparison network under different degrees of occlusion.