Research On Urban Dog Face Detection Algorithm Based On Faster R-CNN

As the number of urban dogs proliferates,a range of dog management issues arise in society.Therefore,the construction of intelligent dog management has become one of the priorities in the construction of civilised cities.Accurate identification of dogs is the key to intelligent dog management,and dog facial detection is a core part of this.In this thesis,the classical algorithm of target detection,Faster R-CNN,is used as the basis for the research,combined with deformable convolutional networks,residual networks,feature pyramids and other techniques to carry out the research of dog face detection.In this thesis,The research is mainly used to assist the relevant departments to solve the health and security problems caused by urban residents keeping dogs,to simplify the processes of dog quarantine,vaccination and regionalised management,and to improve the level of information,intelligence and systematic management of urban dogs.The main contents of this thesis are as follows.(1)Establishing the dog face detection dataset MDog,with data sources from domestic and international social platforms,corporation,Dog Retention Institute and the open source fine-grained classification dataset Tsinghua Dog.The MDog dataset containing 70,428 images of dog faces was obtained.(2)Proposed an improved dog face algorithm for the backbone network,using the residual network Res Net50,the feature pyramid network FPN and the deformable convolution Dconv,to improve the backbone network VGG16 of the original Faster R-CNN algorithm.Res Net50 achieves deeper network layers to improve the detection accuracy of the algorithm;FPN combined with Res Net50 achieves the fusion of features from different layers to improve the detection capability of multi-scale targets;deformable convolution replaces the standard rectangular convolution to improve the sampling capability of targets with different morphologies.(3)The proposed improved dog face detection algorithm for candidate region pooling uses bilinear interpolation to calculate the eigenvalues corresponding to floating point coordinate values to reduce the feature loss caused by coordinate rounding during candidate region pooling;deformable pooling is used to assist sub-region division to improve the adaptive localization capability for different morphological targets,and bilinear interpolation is used to calculate the eigenvalues corresponding to sub-regions to reduce the feature loss caused by rounding during sub-region division.This thesis designs multiple sets of comparison experiments based on the MDog dataset,records the corresponding experimental results,and analyses the experimental results.The results show that the improvements made in this thesis can effectively improve the detection accuracy.The results of the preliminary coupling with the collaborating companies show that the final proposed algorithm can achieve multi-scale detection and multi-morphology detection of dog faces,which meets the expected detection results.