Group-housed Pig Detection Based On Improved CenterNet Model

Traditional breeding industry is developing in the intelligent fields helped by the AI image technology,which will not only decrease management cost but also risk of animals plague.China began increasing funding to construct smart breeding industry after African swine fever broken out in 2018.In the processing of building of the system,machine vision has been becoming more striking among other method for its intuitiveness and low-cost.Few month should be necessary for forming a usage modeling of pig single target,which are able to realize automation detections.However there were no effective models balanced the real time and accuracy.And the existing models were too bulky,who would face lots of difficulties when setting in the computer nodes(MobileNetv2-FPN-CenterNet,MF-CenterNet),The thesis improved a kind ofNet for larger size of animals,pigs,detection constructed in MF-CenterNet.The basic frame is CenterNet cohesion a MobileNet for features gaining,who decreases the net size and makes the detection speed faster.Lots ofNets are compared and the hyperparameters tuned into the best performing are found by thesis.And the improved MobileNetv2-CenterNet performs better than CenterNet with faster speed and smallNet shape size.The analysis finishes a results that the live pig images,nearly half,are small targets,then a MF-CenterNet are added for fixed it.It realizes that while ensuring the lightweight and real-time performance of the model,it effectively improves the detection accuracy of occluded targets and small targets.We test the model in a videos by frames drawing of 1683 samples,which expend to 6732 samples after enhance.The packed size of our models is about 21 MB,who could be set in edge nodes easily with average ACC of 0.9430 in 69 FPS.Compared with Faster-RCNN,SSD,YOLOv3,YOLOv4 target detection network,the detection accuracy is increased by 6.39%,4.46%,6.01%,2.74%,and the detection speed is increased by 54 FPS,47 FPS,and 45 FPS.The improved light MF-CenterNet model can be the real time and improve the accuracy of group pigs providing a new approach.