Research On Key Technologies Of Dairy Cow Lean Feeding System Base On Deep Learning

The scientific feeding of dairy cows means that information such as their behaviour needs to be available in real time.Dairy precision feeding systems are of great value for improving the efficiency of dairy farms and for research into scientific feeding solutions.Current research on key technologies for data monitoring in dairy precision feeding systems is generally based on wearable devices to monitor animal status and obtain animal behaviour data.Analysis of data with the help of statistical,machine learning and deep learning methods has also been applied to some extent.However,wearable devices need to be customised,worn and maintained at a high cost,depending on the subject being monitored.While individual wearable devices such as these can ensure the accuracy of monitoring to some degree.the feasibility of trying to equip each individual cow with a high-cost wearable device is not high under current conditions.Due to the cost of equipment,many of the methods used to monitor cow behaviour on dairy farms are not widely available.Furthermore,dairy precision feeding systems need to be supported by cow condition monitoring data in order to scientifically develop feeding programmes.The correct identification of individual cows is a prerequisite for the effective use of cow monitoring data.Single-target cow identification methods based on computer vision are often difficult to apply due to the uncertainty of target occlusion and target pose in practical multi-objective scenarios.Therefore,it is important to develop a low-cost,easy-to-apply and real-time multi-objective cow behaviour monitoring model for fine dairy farming.For the cow behaviour monitoring problem,a multi-target cow feeding behaviour recognition method is proposed,based on the YOLOV3 algorithm,to achieve cow feeding behaviour monitoring by classifying cows in the barn into three types of targets based on target differences,in order to monitor the feeding behaviour of multiple cows by a single device.The YOLOV3 algorithm has shortcomings such as high computational cost,high energy consumption and strong equipment dependency,and for this problem,a better sparse sub-network screening method based on the amplitude iterative pruning algorithm was proposed with reference to the lottery hypothesis,resulting in an 87.04% reduction in the number of parameters and a mean average precision(m AP)of 79.9%,an improvement of 4.2% over the original network.The feasibility of reducing the cost of cow behaviour monitoring tasks by means of amplitude iterative pruning techniques is illustrated,and the effectiveness of screening better sparse sub-networks from the cow feeding behaviour recognition model based on the lottery hypothesis is verified,providing a reference for the cost reduction of animal behaviour monitoring tasks.For the cow identification problem,a multi-target cow identification method based on attention transformation network is proposed.The multi-target tracking algorithm is used to obtain the image sequences of cow targets,and the short-time occlusion problem between targets in complex scenes is solved by recognising the image sequences one by one and voting on the recognition results.During the recognition process,attention weights are reasonably reallocated in order to be able to mask the low-level attention noise acting on non-class activation regions.In this dissertation,a Self-attention focusing algorithm is proposed,in which the smaller connections in the attention weight matrix are set to zero,so that these positions are no longer involved in the calculation of the attention score,and the attention weights are reallocated,so that attention is gradually focused on the class-activated regions during the iterative process,and the background part no longer shares the attention weights,thus achieving the purpose of shielding the background noise.The experiments show that the accuracy of the model has been improved by 1.7 percentage points after the implementation of the Self-attention focus algorithm,provided that the model size has been reduced by 34.3%,proving the effectiveness of the Self-attention focus algorithm.This dissertation also innovatively reuses the depth epistasis features extracted by the re-identification module of the multi-objective tracking algorithm as the input of the Self-attention transform network,enabling the model to capture the long-range dependencies in images while still having the inductive bias characteristics of a CNN.While saving computational resources,this enables the model to achieve a recognition accuracy of 88.54%,an improvement of 3.1% over the original model,and enhances the robustness of the model to changes in target morphology.The feasibility of solving short-time target occlusion by target tracking is demonstrated,and the effectiveness of adding convolutional features to the Self-attention transform network is verified.Methodological support is provided for the integration of data collection modules in dairy precision feeding systems and for further optimisation of recognition accuracy.