.png){kind=icon}
.png){kind=icon}
Return
Cattle incremental learning identification method based on phased dynamic expansion network
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.128019.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In recent years, as ranching has developed toward large-scale, modern, and efficient practices, simple and accurate cattle identification techniques have become increasingly important. Cattle identification serves as a foundation for intelligent livestock management and also plays a critical role in related financial sectors, such as livestock insurance. However, the characteristics of neural network closedness lead to the need to spend a lot of computational cost and time cost to retrain after a new batch of cattle acquisition, which limits the application of neural network. By analysis of data characteristics and network similarity, this article proposes Phased Dynamic Expansion (PDE) method. Only part of old data and new data are used to achieve the accuracy of Joint-train, saving a lot of computing resources and time costs. The PDE divides the network in terms of phases, and applies the idea of Gradient Boost to the deeper phased network and classifiers, retaining the old deep network to preserve old knowledge and using the extended deep phased network to learn knowledge specific to the new category. For the problem of imbalance in the number of samples between the old and the new, samples are weighted using the number of samples. Two auxiliary classifiers are added to classify the new and old categories respectively. Finally, knowledge distillation is used to avoid excessive parameters caused by network expansion and reduce the requirements for hardware. The difference between the PDE results and those of the Joint-train is 1.31%, 0.98%, and 0.55% for experiments with increments of 10, 20, and 50 using the homemade dataset, while validation on the three publicly available datasets achieves a performance close to that of the Joint-train.
Keywords:
Cattle
Incremental Learning
Phased Dynamic Expansion
Gradient Boost
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W