Research On Face Recognition Of Cattle Individuals In Complex Background Based On Capsule Network

The demand for meat,milk,and dairy products is expanding as people’s quality of living rises.The focus of agricultural policy has shifted from production and output value to quality and hygiene simultaneously.Food safety is receiving increased attention and importance from the government,society,and consumers.Individual identification of cattle plays a vital role in food traceability,identification,production management,immunization,illness prevention,and livestock ownership distribution.Invasive techniques of individual cattle identification,such as ear tags,have many disadvantages.For example,the equipment is readily lost,destroyed,copied,and altered.In contrast,cattle face as a visible biometric characteristic has essential biological information.Under non-intrusive conditions,an individual identification can be achieved by cattle face.Individual cattle face recognition belongs to the area of instance-level recognition challenges.The variety between individual cattle faces is not significant.However,due to background,viewpoint,illumination,deformation,and other factors,the diversity of features between different samples of the same subject is significant.These are the main difficulties and challenges facing individual cattle face recognition worldwide.Based on the preceding,this thesis addresses the challenge of individual cattle face recognition in various complex scenarios using biometrics,computer vision,and Deep Learning,specifically the relationship between Deep Learning models and the quantity of individual samples.The primary research findings and conclusions of this work are as follows:1)In order to gather high quality and consistent cattle face image data,a standardized data acquisition strategy for individual cattle faces in complex circumstances is designed.Based on the actual scenario of large-scale farming and major technological features of data acquisition,guidelines for individual cattle face sample collection and data capture needs are proposed after investigations on farms.Based on the standards and conditions,individual cattle data is collected,and the original cattle video and image datasets are made.The Res SSD algorithm is proposed to complete the extraction of cattle face photos from the original cattle datasets.Summarize the preceding study and construct a cattle individual facial data acquisition technique.2)There are few large-scale public face datasets of cattle individuals.In this study,we construct 264 individual cattle face image datasets.The dataset includes 143,751 files altogether,including individual cattle face pictures,individual cattle face annotation files,and individual cattle face images.The datasets can be used for image recognition and target detection tasks.This research suggests the M-CVAE algorithm for generating facial images of specified individual cattle in order to compensate for the imbalance between the sample sizes of each category.3)The C-LBP Caps Net algorithm is proposed for individual cattle face image recognition in the case of multiple individual samples.First,a C-LBP fusion feature extractor is proposed by combining convolutional features and local binary pattern texture features.Then,the classical Capsule Network structure is improved by introducing a self-attention mechanism and adding an intermediate capsule layer to enhance the feature extraction capability while improving the capsule utilization.The C-LBP fusion feature extractor and the improved Capsule Network are used to propose the C-LBP Caps Net algorithm.C-LBP Caps Net is used in comparison experiments with other networks on individual bovine face datasets.The experiments demonstrate that C-LBP Caps Net achieves 99.12% accuracy and 98.84% F1 value in individual cattle face recognition.C-LBP Caps Net also shows good performance and robustness compared with other networks when new bit-pose data is added during the training process.4)The Siamese DB Caps Net algorithm is proposed for image recognition in the case of small individual samples.To increase CNN’s ability to extract features while preserving the spatial information of those features,dense modules are first fused to improve the structure of the traditional capsule network.Then,to obtain the bivariate features,the Siamese Network structure is employed to extract the picture pair features.Correlation analysis of bivariate features is used to substitute the distance metric of classical Siamese Network for image identification,and it is based on the invariance feature of the bit-pose vector.Comparative experiments are conducted with other networks on a small sample of individual bovine face datasets.The experiments demonstrate that Siamese DB Caps Net achieves 93.00% accuracy and 93.54% F1 value.5)A zero-sample learning method for individual face recognition of cattle is proposed for image recognition in the case of individuals without training samples.According to the Siamese DB Caps Net Network,bit-posture vectors can extract cattle face features,and similar bit-posture vector features can be mapped from cattle face that does not participate in training.The optimal model trained with Siamese DB Caps Net is tested on a dataset of 30 unfamiliar cattle faces that did not participate in the training.The individual cattle face recognition rate reaches 86.92%,demonstrating that the Siamese DB Caps Net model can perform zero-sample face recognition of unfamiliar cattle.