| Abnormal behaviors of pigs are often related to their environment,nutritional status,psychological status and other characteristics.Timely and accurate observation of abnormal behaviors of pigs is of great significance to the maintenance of pig welfare.Pig activity video contains a lot of its survival state.Conventional pig information extraction methods,such as data entry of large-scale pig farm production administrators,manual collection of free range farmers,and wearing sensors for pigs,are slow in collection speed and incomplete in data classification.At the same time,physical operation on pigs may cause stress reaction and chain reaction of surrounding pigs.Machine vision technology can quickly collect pig videos and extract image features on the basis of not touching the pig’s body.It can classify and recognize the behaviors of pigs with large range of activities,such as arching,biting,attacking,jumping,etc.This method can not only avoid the stress reaction of pigs caused by contact,improve the welfare of pigs,but also have a positive impact on the economic benefits of pig farms.However,because most pigs are in motion,it is often not effective to recognize pig behavior directly.Therefore,based on the data before and after video segmentation,this paper uses neural network for recognition,and then compares the results of the two methods to draw a conclusion.Saliency geodesic video object segmentation method uses spatial edge and temporal edge to detect saliency,which can extract object contour accurately.However,video object segmentation based on saliency geodesic may lose part of foreground information,and can not reduce background noise well.Clustering algorithm can induce similar classes,so as to remove the irregular noise information.Based on the above problems,this paper first combines the salient geodesic video target segmentation method with the density peak clustering module to construct a new segmentation method,in order to filter out the noise around the target as much as possible.At the same time,the binary matrix generated from the segmented image is dot multiplied with the original image matrix to retain the texture and color features of the pig itself to a certain extent.Finally,the deep learning method is used to recognize the abnormal behaviors of pigs,such as wall climbing,fence biting,attacking,arching and normal behaviors.The influence of image segmentation on the accuracy of pig behavior recognition is discussed in the form of chart.The main conclusions are as follows:(1)The average accuracy(m PA)of this method is 94.65% and 92.91% respectively,which is verified on Davis and PIGI data sets.Compared with the original algorithm,the accuracy is improved by nearly10%,which shows that this algorithm can reduce the interference of background noise and has good segmentation effect.Compared with other algorithms in recent years,this algorithm combines the advantages of temporal information,spatial information and clustering algorithm,which can eliminate the background redundant information and improve the target integrity.(2)The results of PIGII abnormal behavior recognition show that the average recognition accuracy of Alexnet,Incisionv3,Xceptionnet,Mobilenet and Res Net series networks for pigs is 91.03%,97.99%,97.74%,95.83%,90.57%,94.31% and 95.45% respectively,and the classification accuracy is better than that before segmentation.In this paper,the segmentation of pig behavior data can make the neural network more stable to obtain the characteristic information of the data,and greatly improve the classification results,so as to provide targeted and timely technical support for the intelligent breeding and management of pigs. |
PDF Full Text Request