| With the promotion of large-scale breeding of the dairy goat,the precise monitoring and efficient management of the dairy goat farm has become an urgent problem to be solved.Through the semantic segmentation of dairy farm images,observing the farm environment and the posture and position of the dairy goat,the behavior of the dairy goat can be monitored in real time,and the health problems of the dairy goat can be discovered in time.This paper uses the dairy goat farm images in animal husbandry teaching experimental base of Northwest A & F University as the research object,constructs the semantic segmentation dataset of the goat farm images,improves the decoding network of DeepLabV3+,optimizes the ASPP structure,combines the channel attention mechanism,and uses extracted image features to realize the semantic segmentation of dairy goat farm images.The main research contents and conclusions of this paper are as follows:(1)Semantic segmentation dataset construction of dairy goat farm images.In view of the insufficient dataset available for semantic segmentation of dairy farm images,the data is collected by network camera and filtered manually,and then the ground truth of semantic segmentation is obtained by labeling the dataset with Labelme software.Finally,the dataset is enhanced by flipping,rotating and scaling,and the training dataset and testing dataset are divided,laying a data foundation for subsequent research.(2)Semantic segmentation of dairy goat farm images based on improved DeepLabV3+network.Aiming at the problem that the edges of the segmentation result of the semantic segmentation model are not precision,an improved DeepLabV3+ semantic segmentation method for dairy goat farm images,DP-DeepLabV3+ is proposed.The feature fusion times are increased in the decoding network,and the feature information of encoding network is fully used,and compare it with a variety of classic semantic segmentation models on the dairy goat farm images semantic segmentation dataset.The experimental results show that the PA and MIo U of DP-DeepLabV3+ network are 93.75% and 75.19% respectively,which has a high boundary segmentation precision.(3)Semantic segmentation of dairy goat farm images based on channel attention.Aiming at the fuzzy and missing segmentation of small-scale target in DP-DeepLabV3+,a semantic segmentation model SE-DeepLabV3+ of dairy farm images based on channel attention is proposed.On the basis of DP-DeepLabV3+,the ASPP module is optimized to improve the utilization of feature information and the channel attention module is added to highlight key features,which improves the segmentation accuracy of the model.In order to further improve the performance of the model,the model is optimized by using extracted image features as network input.It is found that the original image combines edge features as network input performs best,which can effectively improve the segmentation performance of small-scale targets.The experimental results show that SE-DeepLabV3+ has a good segmentation performance,with PA value of 96.34% and a MIo U value of 78.23% on the dairy goat farm images dataset.In conclusion,this paper realizes the semantic segmentation of dairy goat farm images based on DeepLabV3+ network,which plays an important role in promoting the large-scale breeding of dairy goat and improving the intelligent management level of dairy goat breeding. |
