| In large-scale pig breeding production,there are few material level detection equipment in the feed tower.The traditional detection methods are mainly manual detection and empirical judgment.With the emergence of smart pig farming,deep learning has been well applied in animal husbandry.In order to meet the refined management,improve the deficiencies in material level detection,and achieve more efficient and accurate animal husbandry,this paper designs the detection system,and uses the semantic segmentation model combined with the self-built experimental platform to complete the analysis of the feed allowance in the silo detection.In the task of material level image segmentation,by comparing the performance indicators of different algorithms,U-Net is more in line with the actual needs and shows good segmentation ability.Therefore,the material level detection method based on U-Net has great research and application value.This paper studies the U-Net semantic segmentation model,optimizes its structure in combination with the real scene,and improves the detection accuracy and accuracy.The main research contents of this paper are as follows:(1)Study the working principle and applicable characteristics of semantic segmentation algorithms in deep learning,understand the calculation principles of various classic network models,and choose U-Net as the semantic segmentation model for this research by comparing their advantages and disadvantages.Conduct in-depth research on UNet’s network layers and their connections,and use this model reasonably in experiments and analysis.(2)Through experiments,the U-Net semantic segmentation model is reasonably adjusted to make it a two-class model to reduce the complexity of the target category.In the network structure,the residual structure and spatial attention mechanism are introduced to improve the network’s recognition ability and Segmentation effect,optimize the U-Net semantic segmentation model,and finally perform image processing on the output image to detect the remaining material capacity in the silo.(3)Under the same environment configuration,the improved U-Net network model and the original U-Net network model were used to test the same data set respectively.The category average pixel accuracy also increased by 0.83%,the F1 score increased from 0.942 to 0.951,and the material level detection accuracy increased from 90% to 94%.It shows that the material level detection method based on improved U-Net has certain feasibility and can meet the basic requirements of detection.(4)Build an experimental platform by yourself,consider various influencing factors in the experiment,and complete the design of the detection system.The residual material value display function is added in the material level detection process,which can intuitively and conveniently observe the residual material capacity in the current material tower.The degree of automation is higher,which effectively reduces labor costs and improves production efficiency.This research reasonably combines deep learning with material level detection,which provides a certain theoretical basis and technical support for the detection of rest feed in the large-scale breeding production and the development of modern pig scale breeding. |
