Research on the posture recognition method of suckling sow based on deep learning

The posture of the sow is one of the important evaluation indicators of its maternal behavior, so it has received extensive attention from breeding and breeding management. At present, the recognition of the sow's posture is mostly based on machine vision and three-axis acceleration sensors. However, when using machine vision technology to capture individual posture images of pigs in the rearing state, the efficiency of sow posture recognition is low due to factors such as light intensity, complexity of the environment in the pen, and location of the pigs. In this paper, 18 postpartum Dumin (Dulac × Min pigs) sow postures were collected for 8 days of video data, and the sow posture recognition model was established based on the Dual-Resnet34 method. Offline and online data enhancement methods are used to enhance the captured image and efficiently decouple the posture features to improve the efficiency of model training. A dual-channel residual network combined with a large convolution kernel is proposed to linearly determine the pig features, which improves the accuracy of the model. The results show that the model proposed in this paper has a positive rate of 99.68%, 99.99%, 99.06%,99.5%, and 99.99% for the 5 types of postures of lactating sows on the test set. And the overall recognition rate of various gestures of pigs reaches 99.5%, which is an increase of 0.42% compared to the single-channel Resnet34. Therefore, this method improves the recognition accuracy while ensuring the recognition efficiency, which can provide a technical reference for the automatic recognition of sow behavior.