基于逐元素动态融合与自适应损失函数的牛脸识别算法

Cow Face Recognition Algorithm Based on Element-by-element Dynamic Fusion and Adaptive Loss Function

摘要:

针对牛只识别任务中存在特征学习敏感性差、牧场背景复杂、冗余信息干扰严重及高相似度个体区分困难的问题，提出一种基于逐元素动态融合与自适应损失函数相结合的牛只面部识别方法。采用StarNet架构对特征提取网络进行重新设计，构建一种融合多层次特征的高效网络，用于提炼出关键的视觉信息，增强模型对细微差异的敏感性;提出一种逐元素特征融合模块，对低维特征进行筛选，对不同特征图进行加权融合，确保重要特征得到保留，而无关或冗余的特征则被抑制和去除;设计一种动态自适应损失函数，将ArcFace Loss的角度边界进行自适应调整，从而增强模型对高相似度样本的区分能力，且能够自适应地平衡不同类别之间的特征分布，显著提升整体识别性能。在自建数据集和公开的300头牛只数据集上进行算法验证，准确率分别达到88.42%和86.67%，表明了算法的有效性和优越性。将算法移植到嵌入式平台上进行性能测试，运行速度达到15 f/s。相比其他算法，本文算法在牛脸识别上有更好的识别效果。

Abstract:

Aiming at the problems of poor sensitivity of feature learning, serious interference of complex redundant information in pasture background and difficulty in distinguishing high similarity individuals in cattle recognition task, a cattle face recognition method based on element-by-element dynamic fusion and adaptive loss function was proposed. Firstly, the StarNet architecture was used to redesign the feature extraction network, and an efficient network integrating multi-level features was constructed to extract key visual information and enhance the sensitivity of the model to subtle differences. Secondly, an element-by-element feature fusion module was proposed to screen low-dimensional features and perform weighted fusion on different feature maps to ensure that important features are retained, while irrelevant or redundant features were suppressed and removed. Finally, a dynamic adaptive ArcFace Loss was designed to adaptively adjust the angle boundary of ArcFace Loss, so as to enhance the model's ability to distinguish high similarity samples, and adaptively balance the feature distribution between different categories, which significantly improved the overall recognition performance. The algorithm was verified on the self-built data set and the public data set of 300 cattle. The experimental accuracy reached 88.42% and 86.67%, respectively, indicating the effectiveness and superiority of the algorithm. The running speed reached 15 f/s, and it was applied to complex environment testing. Both of them achieved high accuracy, which proved that the algorithm had high stability and robustness. Compared with other algorithms, the proposed algorithm had better recognition effect on cattle face recognition.