Research On Visual Inspection And Body Size Assessment Algorithm Of Beef Cattle

Deep learning and computer vision are very important for the development of intelligent animal husbandry,but at present,basic tools such as models and datasets in animal husbandry scenes are still insufficient.In the process of beef cattle breeding,farmers need to regularly monitor the weight of beef cattle.The traditional weighing methods have cumbersome procedures and cost a lot of manpower and material resources.The body weight of beef cattle is directly proportional to the body size.Therefore,in this thesis a non-contact method to evaluate the body size of beef cattle is proposed to replace the traditional body weight measurement evaluation method.The main work is as follows:First,for beef cattle body size evaluation,two relevant data sets are constructed,namely cattle object detection data set with ROBB annotation and segmentation data set of single cattle from ROBB data set.The arbitrary orientation target detection data set is a single category,with multiple instances per frame,and has a direction aligned with the body orientation and a non-rigid target.This data set has 5587 images.The image segmentation data set is used to segment a single cow target based on the target detection frame with arbitrary orientation.This data set has 520 images.Second,it is necessary to master and monitor the trend of beef cattle in the process of beef cattle body size assessment.To solve this problem,in this thesis a beef cattle recognition framework is presented,which based on single-stage target detection model with arbitrary orientation target detection and overhead view.Firstly,the orientation angle parameter is constructed based on the horizontal frame.The random orientation detection model is responsible for predicting and framing complete beef cattle.Secondly,considering that slight changes in the orientation angle will lead to incomplete beef cattle in the frame and affect the evaluation of beef cattle’s body shape,the accuracy of the orientation angle needs to meet the task requirements.To solve this problem,two orientation sensitive IOU algorithms are designed,namely the Close IOU algorithm and the Minor Angle algorithm,which add the angle factor to the candidate box screening mechanism.Then,the loss functions Direct COSLoss and Minor COSLoss related to different angles are designed to deal with the periodicity of upward orientation and improve the orientation angle.Finally,the Close IOU and Minor Angle algorithms and the loss functions Minor COSLoss and Direct COSLoss are combined into the arbitrary orientation target detection framework.The experimental results show that the performance of the combination of IOU algorithm and loss function proposed in this thesis is better than that of baseline IOU algorithm and MSE angle loss on the classical evaluation matrix and orientation constraint evaluation matrix.Third,in order to evaluate the beef cattle’s body shape,Deeplabv3+ semantic segmentation model is used to segment the ROBB box selected beef cattle image,and the number of pixels is counted to complete the evaluation of beef cattle’s body shape.The experimental results of cattle segmentation and body recognition are presented and analyzed.This work is a pioneering exploration of the detection of non-rigid oriented objects(the direction is [0,2 π).It will provide a solution to similar problems in the detection of single category,non-rigid,and oriented objects in animal husbandry and manufacturing.