Research On The Detection Of Broilers’ Pathological State Based On Machine Learning

Efficient and healthy broiler farming technology is an important guarantee in the process of intensive broiler farming.Automatic detection of the pathological status of individual broilers or small group of broilers through machine learning enables early detection of disease at all stages of growth,and can reduce manual patrols,provide technical support for fully automated and unmanned healthy farming.In this study,machine learning techniques such as target identification and detection were used as theoretical research methods to investigate how to quickly and accurately detect the pathological status of broilers in a complex background,using healthy and diseased white feathered broilers and yellow feathered broilers as study subjects.The following methods were used:1)Establishment of a database of broiler pathology tests.The information was used to find images of typical characteristics of broiler chickens for various diseases,including crown,beak,eyes,feathers,legs and claws.Borrowing from HSV Color Spatial Transformation and CLAHE.Sharpen the background by changing the saturation and tone of the image through HSV color space,and enhance image brightness and detail by CLAHE.Finally,building a database of pre-processed broiler images by normalizing images.2)Constructing a broiler pathological status assay classifier.Use the directional features in Haar to collect image details,and then use the Ada Boost classification algorithm to train the cascade classifier,apply the generated classifier to the broiler image,obtain the target broiler return frame information in the original image,and realize the rapid positioning of the broiler under complex background with recognition.3)An enhanced Squeeze Net architecture is proposed to train broiler pathological state detection models.Based on data enhancement,the Squeeze Net algorithm is used to further optimize the basic feature extraction,multi-layered complex feature extraction,weight learning,and so on.Training enhanced broiler dataset with the Tensorflow deep learning framework improves detection accuracy.In this experiment,200 images of broiler chickens were enhanced,and after 10 iterations,2000 images of broiler chickens were obtained,which were divided into 1000 white feathered broiler chickens and 1000 yellow feathered broiler chickens,each with half of healthy and diseased images.It has been verified that the average accuracy AP⁵⁰and the recall rate AR of healthy white feather broiler and diseased white feather broiler are 0.892,0.511 and 0.887,0.527 by using the improved Squeeze Net model;the average accuracy AP⁵⁰and recall AR for healthy yellow feathered broilers and diseased yellow feathered broilers were 0.816,0.457and 0.853,0.492;respectively,and the m AP of the final model was 0.862 with a correct rate of 94.8%.Whereas,the m AP of Squeeze Net model was 0.788 with a correct rate of 86.7%.The study shows that the enhanced Squeeze Net model can quickly detect the status of diseased broilers with better performance than the original Squeeze Net model based on machine learning.The model can be generalized to other domestic animals such as pigs,cattle and sheep for pathological status detection.