Pig Target Based On Deep Learning Inspection And Status Analysis

The living situation and disease prevention of commercial pig in pig farm mainly depend on human eyes.This visual inspection is labor intensive,time consuming and suffers from the problem of individual experience and varying environment.It has been demonstrated that if the pig keeps behaviors such as standing with rising head or laying for longer time,the pig is likely tired or ill.It needs a further health check and take a proper treatment.If the pig keeps standing with head rising,the pig could be stimulated and disturbed leading to panic or fear by external environment.Therefore,it is necessary to examine the environment of the pig farm.In order to efficiently observing pig growth state,the method of video monitoring has been applied in the field of pig farm.This method not only monitoring the living situation of pig in real-time,but also highly reducing human labor and production cost.The pig recognition is very important for pig behavior analysis and digital breeding.For obtaining detailed pig growth state and discovering potential disease,this work firstly detected the pig object using deep learning technique.In this step,a kind of improved ResNet model was proposed in this paper.The developed model improved the detection rate with less net parameters.The developed model captured the features of the pig applying across layer connection.And the ability of feature expression was improved by adding new residual module.The number of layers were reduced to lower the net complexity.In general,the ResNet frame was developed by reducing the number of convolution layers,constructing different type of residual module and adding the number of convolution kernels of certain convolution.The developed model efficiently improved the ability of feature description for ResNet with less deep.The training process was easy while the classification accuracy was improved in the developed model.The experimental result illustrated that the developed model still has the ability of feature expression under the condition of less deep,and the detection rate was improved.The successful object detection is important for pig feature extraction and pig posture recognition.The result of detection directly influences the detection rate of the pig posture.In this paper,image filtering,image enhancement,image edge extraction and classification task were studied.A kind of improved bilateral filter was proposed for extracting pig feature.The experimental result illustrated that the method could extract the pig tour.The specific research contents and results are as follows:1)The related theory of deep learning and the animal posture recognition are explored in this paper towards to developing a new method for commercial pig detection.The proposed method is less influenced by environment and this method can analyze pig behavior for pig farm.2)After analyzing individual pig feature,the experiments were performed applying CNN of AlexNet^GoogLeNet and ResNet.A kind of improved CNN for pig detection was presented and several different data sets were used to train and test the improved model.The experimental result illustrated that the detection rate and the performance of the model is better when using large data set.The training accuracy is 94.2%,96.7%and 97.7%when using AlexNet?GoogLeNet and ResNet model.The testing accuracy of 90.4%,93.1%and 95.3%was achieved when using these three models.When using the improved model,the training accuracy and testing accuracy reached 98.2%and 96.4%,respectively.Compared with those models,the improved ResNet model is outstanding.Therefore,the improved model can detect individual pig efficiently.3)The outline of pig was analyzed in this work.The method of shape description based on object border was used for descripting the shape of pig which can meet the requirement of real-time detection.For efficiently detecting pig posture,the method of shape fitting was applied in this paper.The general shape of pig was fitted using four ellipses.The experimental result demonstrated that the diversity difference in intra class can be ignored when descripting pig outline using ellipse fitting.4)The algorithms of image filtering,image enhancement and edge extraction were studied in this paper.The method of Canny edge extraction based on a kind of improved bilateral filter was proposed for extracting pig feature.The experiment showed that this method can extract the outline of pig efficiently.5)The recognition of pig situation was explored and a kind of KNN classifier based on template matching with weighted Euclidean distance was proposed for recognizing pig situation.The experiment showed that the proposed classification method obtained higher recognition accuracy with low computational cost.