Cattle Face Detection And Attitude Angle Estimation Based On Deep Learning

As the industry of animal husbandry gradually develops large-scaly and intelligently,the efficient and stable traceability of cattle becomes the key to ensuring the safety of meat products and controlling the spread of diseases.Identity authentication,which is the most effective way to approach traceability,is to bulid database after detecting the cattle's face and extracting features.The effect of the face detection which is a main step of identity authentication determines the accuracy of identity authentication directly.In order to collect the features of various angle of cattle face,the attitude angle of cattle face should be detected,which can take feature collection as a reference.This study mainly aims to the detection of cattle face.Firstly,the basic network of SSD and the aspect ratio of default box are optimized,which improves the detection performance of the model.Then,considering the attitude angle estimation,the cascade model and the DAE(Detection and Angle Estimation)model are constructed based on the improved SSD.The main research contents are as follows: 1.A method of cattle face dataset is built.Firstly,the dataset collects approximately 3,800 images of cattle face including scenes of single and multiple cattle.Then,the annotation file of the dataset conducts necessary parameters such as image size,bounding box position and attitude angle which provides different parameters for different tasks.Finally,the data is conducted to random-crop,distort-color and other operations,which can enhance the diversity of data and ensure the training effect.2.A method of cattle face detection is proposed.Firstly,the basic network and aspect ratio settings of default box are optimized,which improves the model of SSD.Then,MobileNet v2 is used as the basic network of SSD and depth-wise convolution is employed in additional feature extraction layers.Finally,K-Means++ algorithm is used to cluster aspect ratio on the samples to set the aspect ratio of default box.3.For the attitude angle estimation,a cascaded model which is based on improved SSD and MobileNet v2 is proposed.The improved SSD is used to detect cattle face,and the results are sent to MobileNet v2 for attitude angle estimation.The cascaded model can support detection and attitude angle estimation in scenes of multiple cattle and ensure high detection accuracy and low angular error.4.Considering the speed problem of the cascaded model,the DAE model is given based on the improved SSD.In order to ensure the attitude angle can be transmitted successfully and participate in the training,the branch for transmitting the attitude angle is added in the architecture of improved SSD and the steps of training and loss functions of DAE model are improved accordingly.Unlike the cascaded model,the DAE model guarantees high precision and speed in scenes of single and multiple cattle.The simulation results show that the improved SSD is superior to other detection models in terms of accuracy and speed for cattle face detection.For attitude angle estimation,the DAE model has the detection accuracy close to the improved SSD and the lower angle estimation error compared with cascaded model,and can ensure a high speed in different scenes.Over all,the results verify the reliability and effectiveness of the proposed method.