Study On Individual Recognition Of Topview Group-housed Pigs Based On Local Features

Identification of group-housed pigs is the foundation of intelligent individual surveillance and behavior analysis.With the specialization,intensification and standardization of China’s pig breeding sector,the demand for pig quality control,welfare management,and disease prevention is increasing day by the day.In-depth research on group-housed pig identification methods and the development of the intelligent monitoring systems are conducive to the automatic supervision of the pig breeding process,which has significant theoretical significance and great application potential for improving animal welfare and quality control.At the moment,research into pig face recognition based on machine vision has made significant progress,although image collecting of pig faces has several limits due to pig mobility and posture.Based on the Project of National Natural Science Foundation of China,this paper investigated the identification method of individual pigs in natural scenes with moving pigs in the overlooking video in order to discover a more convenient and efficient method and identification model for group-housed pig identification.The following are the paper’s primary innovative research findings:(1)A new identification method based on Gabor transformation and local binary mode was proposed.Considering the limitations of collecting face and side views of pigs,the pig back hair,skin texture,and spots of biological characteristics were evaluated while disregarding individual pig body images as the research object.The multi-scale Gabor feature and local binary pattern texture feature were extracted.For classification,the support vector machine was utilized.The proposed method,which explores the feasibility of local texture features for pig identification from a top view and provides a technical premise and application basis for individual monitoring and behavior analysis of group-housed pigs,can identify the 7 pigs in the pigsty with an average accuracy of 91.86 percent.(2)A multi-scale local difference directional number(MLDDN)pattern is proposed for the identification of group-housed pigs.A multi-scale and robust coding method is used based on the study of the effectiveness of local texture features for pig identification,in order to further improve the ability of feature description and to compensate for the shortage of the local direction number model,which only considers a single scale and does not extract the relationship between different directions of response.To begin,color images of pigs were converted to grayscale images using the most significant bits quantization approach,so that distinct hues could be quantized into more discriminative grayscale values.Second,in order to gain more information,not only the Gabor amplitude response but also the illumination insensitive Gabor phase response are extracted.Then,by computing the main variations in local edge directions,the amplitude and phase responses are coded individually.Finally,the histograms of block-coded images are maximized at different scales to avoid the increase of feature dimension induced by multi scales.This method can accurately identify individual pigs in two pigsties with 7 pigs and 10 pigs,respectively,with 95.71 percent and 89.80 percent accuracy,resulting in a new method for local feature extraction of individual pigs.(3)A Weber texture local descriptor(WTLD)was proposed for the identification of group-housed pigs.The feature dimension is great,and the calculation is huge,because both above method apply the Gabor transform to get multi-scale and multi-direction features.A more compact and efficient local feature description method is utilized to increase the compactness of coding and minimize the feature dimension,based on a study of the inadequate representation of the direction and local structural information of Weber descriptor.To begin,Weber’s law is used to determine the differential excitation,which describes the relative intensity difference between the central pixel and adjacent pixels.After that,using the multidirectional template,the local multidirectional information is extracted,and the primary direction information is coupled with differential excitation for the histogram.Simultaneously,the gray variation of local neighborhood in the main direction and its diagonal directions is determined in ordet to improve the description of local structural information.Finally,the direction and strength of the change are encoded using an adaptive threshold.This model,which as low feature dimension and computational burden,can identify individual pigs with 97.1 percent and 95.0 percent accuracy for two experimental pigsties with 7 pigs and 10 pigs,respectively.The above research findings show that by extracting local features from overlooking pig individual images,good performance in pig identification can be attained.This method can lessen the constraint on pig’s movement and posture to some amount,and it provides a theoretical foundation for the research on group-housed pig identification.With less feature dimension and computational burden,further research on feature expression ability achieves better results.This allows the pig breeding business to minimize the cost of configuring the group-housed pigs identification system and increase the intelligent and accurate breeding level of large-scale pig farms.It as significant academic and practical usefulness in improving the intelligence pig industry and improving animal welfare.