Individual Identification Of Dairy Cows Under Unrestricted Conditions Based On Fusion Of Deep And Traditional Features

Individual cow identification is a prerequisite for intelligent cow behavior analysis and disease detection.The traditional method of individual identification of dairy cows has low efficiency and high cost,which can not meet the needs of modern breeding.Cow individual identification based on machine vision has the characteristics of non-contact,high efficiency and no stress.Researchers at home and abroad have made more achievements in individual identification of cows under the condition of postural restriction.Under non-restrictive conditions,due to the change of camera distance and Angle,the cow pattern in the image will be deformed greatly,which increases the difficulty of identification.The depth features extracted by deep learning technology are complementary to the traditional features integrated with expert knowledge.A method of individual identification of dairy cows based on the fusion of depth and traditional features under unrestricted conditions was proposed.The main research contents and conclusions are as follows:(1)Night image enhancement.The light intensity of the cattle house at night is low,and the quality of the acquired image is poor.This paper uses the improved MSRCP algorithm to enhance the night image.The improved MSRCP image enhancement algorithm not only effectively improves the image quality of low illumination,but also significantly improves the MPA of night image segmentation.(2)Identification of cows in active areas.Three traditional features,Hu moment feature,global pattern feature and local shape feature,were extracted according to the characteristics of dairy cows in active areas,and feature selection was carried out by NCA method.AlexNet network was used to extract the depth characteristics of cows in active areas.The individual identification accuracy of dairy cows in active areas was 93.87%,95.70%and 98.66%using traditional,depth and fusion characteristics respectively.(3)Cow identification in rest area.Compared with the active area,the pattern deformation of cows in the rest area is larger.Two traditional features,Gabor feature and LBP feature,were extracted from the images of cows in the rest area,and PCA was used to reduce the dimension of the combined features,and the best PCA parameters were selected to obtain the final traditional features.The improved AlexNet network was used to extract the depth characteristics of cows in the rest area.The individual identification accuracy of cows in rest area was 90.17%,96.89%and 98.94%,respectively,using traditional,depth and fusion features.(4)Selection of feature fusion method.Depth features are less reliable in local spatial features than traditional features,which involve expert knowledge.The two features can complement each other,and the performance of the deep learning method can be improved through the fusion of features,while the expert knowledge is integrated into the model through the fusion of features.Three fusion methods,eigenvalue summation,eigenvalue maximum and eigenweight summation,were used to fuse the features,and the results showed that the eigenweight summation method achieved the highest identification accuracy of cows.In conclusion,an individual identification experiment was carried out on 52 Holstein cows under unrestricted conditions by using the method of individual identification based on the fusion of depth and traditional features proposed in this paper.The results showed that the accuracy of dairy cows in active and resting areas reached 98.66%and 98.94%,respectively,compared with individual identification using depth features alone,the accuracy increased by 2.96%and 2.05%,respectively.Compared with individual identification using traditional features alone,The accuracy rate increased by 4.79%and 8.77%,respectively.