Research On Instance Segmentation Of Dairy Goat Image Based On Improved Mask R-CNN

Precision breeding is currently the mainstream direction of the development of animal husbandry.In order to improve the accuracy of image segmentation of dairy goats and promote the precision and intelligent development of the breeding industry,this paper takes the images of dairy goat obtained in the sheep farm as the research object,and realizes the instance segmentation of dairy goat image based on the improved Mask R-CNN algorithm.The main research contents and conclusions of this article are as follows:(1)Construction of instance segmentation data set for dairy goats.In order to solve the problem of the lack of instance segmentation of the public data set of dairy goats,a remote high-definition camera was installed in the dairy goat farm to obtain indoor and outdoor dairy goat monitoring videos.First,extract the key frames of the video,and manually select the high-definition dairy goat image as the original instance segmentation data set.Then,the method of geometric transformation,color enhancement and noise addition is used to realize the expansion of the data set.Finally,Labelme software was used for image annotation,and an ideal self-made dairy goat instance segmentation data set was obtained.(2)Instance segmentation of dairy goat image based on feature weighted.In order to solve the problem of insufficient use of feature information by Mask R-CNN model,a feature weighted instance segmentation model of dairy goat image FM-Mask R-CNN is proposed.FM-Mask R-CNN adds the SE module in SENet on the basis of Mask R-CNN,learns the importance of feature channels through feature weighted,improves the weight of effective features for dairy goat segmentation,suppress other unimportant features and enhance the feature extraction ability of the model.The experimental results show that FM-Mask R-CNN has higher instance segmentation performance.Compared with the original Mask R-CNN,the average accuracy of segmentation on the self-made dairy goat dataset is increased by 2.83%.(3)Instance segmentation of dairy goat image based on Adaptive NMS and Repulsion Loss.In order to solve the problem of low accuracy and poor robustness of Mask R-CNN instance segmentation model in complex scenes,this paper proposes an improved Mask RCNN algorithm based on Adaptive NMS and Repulsion Loss.First,replace the original NMS algorithm with an Adaptive NMS algorithm that is more suitable for complex scenes,and set the optimal threshold according to the density of the scene.Introduce Repulsion Loss,which weakens the influence of occlusion,to further enhance the Mask R-CNN algorithm’s adaptability to occlude instance targets,and effectively reduce the missed detection and false detection caused by mutual occlusion of dairy goats in complex scenes.Finally,a comparative experiment was completed on the self-made dairy goat data set.Experimental results show that the average segmentation accuracy of the improved Mask R-CNN algorithm based on Adaptive NMS and Repulsion Loss on the dairy goat data set reaches 68.71%,which is 4.59%higher than the original Mask R-CNN.The comparison between the method in this paper and other instance segmentation algorithms also proves that the improved Mask R-CNN has obvious advantages in solving the issue of occlusion in complex scenes,and has better segmentation performance.