Algorithm for Lightweight Cattle Face Identification Based on Improved MobileNet with Dynamic Pruning

Aiming at the problem that most of the cattle individual recognition models based on CNN, Transformer, etc. in the actual ranch require a large number of parameters and cannot be effectively deployed in edge devices with limited computational and reserve resources, in this paper, we improve the backbone of the lightweight MobileNet network and incorporate dynamic pruning based on self-supervised mask prediction, and propose a lightweight cattle face recognition algorithm (Improved MobileNet with Dynamic Pruning, IMP-MobileNet-DP). The algorithm uses a self-supervised approach to learn dynamic routing to help the network automatically learn which channels to activate, by combining the prediction header and heat map quality to generate binary masks for prediction and introducing a dynamic gating mechanism to dynamically prune the improved MobileNet network to maintain the accuracy of cattle identification while significantly reducing the computational complexity of the model and significantly improving the inference speed of the model. Experiments show that the algorithm in this paper achieves a recognition rate of 98.23 %, the number of model parameters is $\mathbf{3. 9 3 M}$, the Flops parameter is only $\mathbf{4 3. 6 0 M}$, and the pruning rate is as high as 91.54 %, which is better than the lightweight MobileNet series network in terms of resource consumption, and it meets the requirement of effective deployment of the cattle recognition model in the real ranch.