Shufflenet-Triplet: A Lightweight Re-Identification Network for Dairy Cows In Natural Scenes

Abstract

Individual identification of dairy cows is the basis of accurate animal husbandry. It also provides technical support for cow breeding, health information acquisition, and dairy product traceability. However, poor portability caused by repeated training for newly added individuals and large number of model parameters is an obstacle to the practicability of current algorithms. The idea of searching in RE-identification can avoid repeated training. In this research, a RE-identification network "ShuffleNet-Triplet" was proposed for individual identification of dairy cows meanwhile avoid repeated training. ShuffleNet v2 was adopted for feature extraction to reduce the parameters. The Triplet Loss and Cross Entropy Loss functions were organically combined to enhance the network's distinguish ability to similar individuals. BNNeck was adopted to reduce the conflict between two loss functions and proved effectively improved network's identification ability. The accuracy rate of the model in the training set was 82.93%. Without retraining the network, the CMC-1 and mAP indicators of the RE-identification network were 94.12% and 73.30% respectively for new dairy cattle individuals. More importantly, the model size was only 5.72 M. It was decreased by 180.28 M compared with the maximum model in the comparison experiment. Compared with the single adoption of the loss function, the proposed model’s mAP was increased by 2.86% and 1.18%, respectively. BNNeck reduced the conflict of the loss function and increased by 6.88% in mAP. The model effectively overcome interference factors such as multi-individual presence, fences, and different postures. All experiments proved the practical value of the proposed model.

Keywords: Individual RE-identification, Dairy cow, ShuffleNet, Triplet loss, BNNeck