Study Of Simmental Cattle Face Dataset

In recent years,with the introduction of national policies,adhere to the "science and technology of cattle",the development of beef cattle breeding industry from the grassland free-range is being transformed and upgraded to the intelligent direction,and the individual identification technology of beef cattle has become an important basis for modernization and refinement of breeding.As an important part of recognition technology,the study of dataset is relatively backward compared with the study of algorithm.Therefore,this thesis constructs a Simmental small-sample cattle face dataset based on cattle facial features,extends this small-sample dataset,and applies image fusion techniques to enhance the cattle face features,and finally conducts a comprehensive evaluation of the constructed dataset.The specific work is as follows:The Simmental small sample cattle face image dataset was constructed first.A total of 103 cattle video data were collected from July to August 2018 and August to September 2020 at Shengyuan beef cattle farm in Hebei Province and Kulun Banner in Tongliao City,Inner Mongolia.After splitting the captured video data into frames,key point interception and image filtering,the final dataset of 10,239 Simmental cattle facial images of 103 cattle was obtained.Secondly,the small sample dataset is extended based on DCGAN network from the initial 10239 cattle face images to20539 cattle face images.The generated images are also evaluated by FID,GAN-train and GAN-test metrics,which demonstrate the authenticity and diversity of DCGAN-generated images for the extension of the cattle face dataset.On this basis,the recognition accuracy under the traditional data expansion method and DCGAN data expansion method are studied comparatively.It is proved that the recognition accuracy of DCGAN is higher than that of the traditional data expansion method.Thirdly,through the individual recognition accuracy of 103 cattle and the production characteristics of the dataset,combined with the image histogram analysis,it was found that there were individual cattle in the dataset are affected by high illumination,resulting in the whitening of the cattle face images,and some cattle face features were lost,making the individual recognition accuracy low.For the above problem,the MEF-GAN model is built for the enhancement of cattle face features.Specifically,a self-attention mechanism is added to the GAN to form a self-attention module,a local detail block and a merge block to fuse new images,and the fused images are evaluated.The experimental results show that some lost cattle facial features due to high light illumination were restored and enhanced.Finally,the Simmental cattle face dataset produced in this paper was validated on several commonly used classical network models VGG16,Goog Le Net and Res Net50,and compared with other kinds of cattle face dataset and other feature dataset of cattle,which proved that the dataset in this paper has good generality and practical application significance.