Research On Behavior Of Yellow-feather Broiler Based On Visual Technology And Deep Learning

The yellow-feather broiler is common poultry in our lives,due to the characteristic of fluffy feathers and dense aggregation,it is difficult for us to identify and determine the location of the yellow-feather broiler accurately in some identification process.The largescale breeding model makes it difficult for the caregiver to find that the yellow-feather broiler is in a sub-healthy state in time,which makes the yellow-feather broiler miss the best treatment period.It leads to an increase in the broiler flock’s death rate and affects the economic benefits of breeding enterprises.In response to the above problems,this thesis took yellow-feather broilers with an age of 50 days as the research object,and collected images of effective interest in complex circumstances through a video monitoring system.The study found that SSD and Faster-RCNN algorithms are difficult to identify yellow-feather broilers accurately.Therefore,this paper proposed the Faster-RCNN-OHEM algorithm to separate the yellow-feather broilers from the complex chamber circumstances and locate the position of the yellow-feather broilers accurately.Due to the density of the broilers and environmental factors,we can analyze the movement patterns of the broilers,and their distribution.The content and conclusions of this thesis is mainly given as follows:(1)In this paper,the image acquisition system and calibration of internal parameters were designed.The environmental monitoring system and image acquisition system were designed according to the solid broiler chamber model built by ourselves.The image data set needed for the experiment was provided for the training model.The Zhang’s calibration method was used to calibrate the internal parameters of the camera to reduce the distortion error brought by the camera and improve the picture quality.(2)In this paper,the yellow-feather broiler images and the making of the image dataset were preprocessed.To solve the problems of a complex environment,uneven light distribution,object occlusion,and a similar color between yellow-feather broilers and the litter,we used the Python language packages to preprocess the yellow-feather broilers images by flipping and panning.At the same time,the yellow-feather image dataset was produced to improve the yellow-feather broiler’s recognition effect.(3)In this paper,the behavior of the yellow-feather broiler was recognized based on the improved Faster-RCNN algorithm.For flocks with the high density of broilers,the SSD and Faster-RCNN algorithms were difficult to identify yellow-feather broilers accurately,we proposed the Faster-RCNN-OHEM algorithm.In order to locate the broilers coordinates position accurately,we trained the yellow-feather broilers sample sets in a complex environment,and established a convolutional neural network model suitable for yellowfeather broilers recognition.As the result,the Faster-RCNN-OHEM algorithm had a recognition accuracy of about 90%for yellow-feather broilers types of eating,drinking,flying and others.(4)In this paper,analysis of breeding welfare based on the behavior of yellow-feather broilers.Based on the improved Faster-RCNN algorithm,the population of yellow-feather broilers at different times and densities was identified and analyzed.It was found that reducing the feeding density of yellow-feather broilers helps to reduce the mutual aggressive behavior between broilers and helps to express the different behaviors of yellow-feather broilers.It is in line with the modern welfare of yellow-feather broilers.The results of this thesis show that:(1)the yellow-feather broiler can be effectively separated from the complex circumstances based on the Faster-RCNN-OHEM algorithm,and the recognition effect of Faster-RCNN-OHEM algorithm is better than that of the SSD algorithm and the Faster-RCNN algorithm.(2)the distribution of yellow-feather broilers with different densities will have a certain impact on space utilization.When the temperature reaches 32℃,the distribution is more dispersed and there will be a slight stress response.This thesis can optimize the level of welfare breeding management,and avoid the occurrence of bad behavior and provide some reference value for disease early warning.