Sow gesture recognition based on semi-implicit iterative algorithm

Abstract:
The posture of the sow is one of the important evaluation indicators of its maternal behavior, which received extensive attention from breeding and breeding management. At present, the methods of sow gesture recognition mainly include machine vision algorithms using machine learning and three-axis acceleration sensors, as well as deep learning algorithms that have been rapidly developed subsequently. However, the deep learning network's process of capturing pig postures is inefficient, that is, the training speed is slow, which is determined by interference factors and the generalization ability of the network. Especially in industrial practice, the training process usually takes days as the unit. This greatly limits the practicality of deep learning method. In this paper, 18 postpartum Du Min (Dulac x Min pigs) sows are collected for 8 days of video data, and the training process based on the deep learning of explicit Euler updates is transformed into semi-implicit Euler updates. The training speed can be increased by 20 times theoretically. And the simulation results show that the application of the method to the neural network improves the training speed by 2 to 3 times. Therefore, this method improves the training speed while ensuring the recognition accuracy, realizes the fast fitting of the deep learning network, and provides a technical reference for the automatic recognition of sow behavior.