Face Detection Of Dogs

Dog face detection is mainly to determine whether a dog's face exists in a given image.If so,the specific position and size of each dog's face in the image are marked.Dog face detection is mainly used for dog face recognition as an alternative means of dog identification such as implanted RFID tags or wearing collars,which are currently complicated in operation and low in social acceptance,or for dogs.Provides focus assistance when shooting.In addition,dog face recognition technology research can be extended to other animal recognition applications.To meet the needs of dog identification applications,this thesis conducts research on dog face detection technologies.The main contributions of this thesis are as follows:(1)The establishment of dog face detection data set.In order to provide a benchmark data set for dog face detection algorithms,an online crawling software was designed and implemented through a data stream-oriented method,which crawled and washed 15396 pictures containing three kinds of dog faces from the Internet;according to LFW data set design rules,the design realizes the screening,classification and labeling of crawling image data,and establishes a "unconstrained" dog face reference data set with uniform sample distribution and conforming to the natural situation.(2)Improvement of YOLOv3 target detection neural network.According to the specific features of the dog's facial image,the YOLOv3 model has been improved in many ways.First,for the characteristics of the Residual Block feature utilization rate of the backbone network in the YOLOv3 network is not high,and the storage space of mobile devices is limited,the Dense Block network structure in Dense Nets with better performance is replaced to improve the detection accuracy of the network and Detect speed and reduce the number of network parameters.Then,for the face of small-sized dogs,by adding an output layer to the output layer,the network can make better use of shallower features and improve the accuracy of network detection.(3)Improved YOLOv3 target detection network verification.The original and impr-oved YOLOv3 networks were compared and tested on the constructed dog face benchmark data set.Through experimental analysis,the YOLOv3₂ detection network obtained after improving the YOLOv3 network can adapt to dog face detection in various environments,and the average detection rate of dog face in various scenarios increases from 85.4% of the original YOLOv3 algorithm At 93.8%,the size of the network model has dropped from 484309 KB to 339303 KB.Feature extraction capabilities have become stronger and more robust.