Research On Detection Methods Of Poultry Abnormal Behaviors Based On Machine Vision

In recent years,the frequent poultry diseases have been hitting the poultry industry hard,causing huge economic loss and threatening human health.Therefore,the poultry diseases have been a major concern among poultry farmers and even for the nation.At present,the monitoring of poultry diseases is mainly based on manual inspections.The breeders conduct preliminary diagnosis of poultry health by observing their posture,feathers,cockscombs,feces and sounds.However,manual inspection is both time-consuming and laborious,with a considerable omission rate during the inspection.Meanwhile,staying for long in the farm may do harm to staffs' health due to a plethora of dust and strong odor.Given the background,this paper proposes a machine vision method to facilitate the disease inspection by identifying abnormal behaviors of poultry to achieve early warning.Currently there are primarily two feeding patterns in both domestic and overseas breeding industry,namely,free-range and cage-raising.In this paper,some machine vision algorithms were devised for abnormal behaviors detection in terms of the two feeding patterns,and the objects are common white broilers and yellow broilers.Since the image segmentation algorithm is the premise and guarantee of the subsequent algorithms,this paper first introduces a variety of poultry image segmentation algorithms applied in different occasions.Thereafter,the paper put forward two tracking algorithms to analyze the movement of poultry.Then the paper studied the abnormal behaviors of free-range and caged broilers from a static perspective.In the study of free-range broilers,the paper firstly studied the posture features of free-range broilers through machine learning methods,and then applied deep learning methods to identify multi-targets and detect their abnormal behaviors.In the study of caged broilers,this research used a depth camera to identify suspected diseased broilers under the manger to detect abnormal behaviors.The research summary are as follows:(1)Multiple image segmentation algorithms were proposed.In addition to the traditional HSV color-space segmentation algorithm and the background-difference segmentation algorithm applied to targeting poultry,this paper also designed a segmentation algorithm according to poultry features based on ellipse model and another one based on both K-means clustering and ellipse model.The latter two are outstanding for whole-chicken segmentation.In order to accommodate complex environmental change,this paper proposes a poultry segmentation algorithm based on depth information,which uses a depth camera to segment poultry in three dimensions.Finally,this paper also proposed an algorithm for segmentation of poultry using deep learning techniques.The paper compared and analyzed three existing deep learning models in the segmentation field and constructed a data set for broiler segmentation in order to train the three models mentioned.(2)Two poultry tracking algorithms were proposed.Target tracking is one of the major difficulties in the field of digital image processing.Target tracking is commonly used in pedestrians and cars while is seldom used in the diagnosis of poultry diseases.This paper put forward a poultry tracking algorithm based on Gaussian particle filter and a multi-target tracking algorithm based on contour features.The test results showed that the former can successfully track poultry targets under poor light conditions and had anti-interference at certain level,which addresses the problem of losing target caused by masking,while the latter is capable of tracking multi-target simultaneously with less amount of calculation and a wide range of applications.(3)Detection methods of abnormal behaviors for free-range broilers based on machine learning was proposed.This algorithm applies to free-range broilers in poultry houses.The algorithm segmented the poultry by applying the combination of the K-means clustering model and ellipse model.By means of thinning method,the topological structure of the broiler was obtained,which revealed the postural differences between healthy and diseased broiler.Eventually by machine learning method the broilers in different health status can be identified and sorted out.After comparison experiments,the algorithm finally selects support vector machine(SVM)to identify and classify broilers.This method has excellent generalization ability and is suitable for small sample learning and disease diagnosis.The SVM model based on polynomial kernel function has a classification accuracy rate of 99.469% in the test set,which is superior to other machine learning algorithms,meeting the accuracy requirements,and minimizing the false detection rate.(4)Detection methods of abnormal behaviors for free-range broilers based on deep learning was proposed.This algorithm applies to free-range broilers in poultry houses.The paper improved the original object detection model based on deep learning,and put forward an object detection model named Improved Feature Fusion Single Shot Multi Box Detector(IFSSD)for disease warning of multiple broilers.The model uses the improved Inception V3 network as the base network,incorporating feature fusion technology,and thereby creating feature pyramid network.The IFSSD model has high accuracy in identification of broiler datasets.The mean average precision(m AP)is 99.7% when the intersection-over-union(Io U)is greater than 0.5,and the m AP is 48.1% when the Io U is greater than 0.9.At the same time,the model can achieve 40 frames per second(fps)under a single NVIDIA 1070 Ti GPU.This method has the advantage that it can simultaneously identifies multiple broilers with a high accuracy rate despite of partly overlapping.(5)Detection methods of abnormal behaviors for caged broilers based on machine vision was proposed.This algorithm applies to caged broilers in poultry houses.After image correction,the key area can be obtained by identifying the manger position,and eventually the suspected diseased broilers with lying prostrate in the key area can be segmented and recognized by using depth camera.The experimental results revealed that the algorithm can achieve 97.80% accuracy and 80.18% recall rate(Io U > 0.5)in white broiler house,and can achieve 79.52% accuracy and 81.07% recall rate(Io U > 0.5)in yellow broiler house.Processed by i5-8500 3.20 GHz CPU,the algorithm speed is 10 fps for images with a resolution of 640×480.This algorithm is marked by its superiority to identify broiler postures(standing up or lying prostrate)without incurring any form of environmental alteration even in the complex cage environment given insufficient light.Then the broilers lying prostrate over certain length of duration would be labelled and further examined.