Research On Dog Breed Identification Algorithm And Application Based On Deep Learning

In recent years,the development of the deep learning technology has strongly promoted the widespread application of image classification tasks in real life.Meanwhile,with the increasingly mature of image classification theory system,this research has made many breakthroughs recently,which is moving towards multi-scenario and large-scale direction.However,most of the current image classification work focus on general classification,such as distinguishing multiple unrelated categories like people and cars.Currently,the research on fine-grained image classification,that is,the identification of multiple different subclasses in the same general category,is still very limited,which has gradually become a research hotspot in current computer vision and its applications.In this paper,we start with fine-grained image classification of dog breed targets,and study the dog breed recognition algorithms and applications based on deep learning.To address the issues of similarities between different dog breeds and large differences between the same breed,the classification algorithm of dog breed images from the perspective of object detection and feature fusion is researched,and the aim is to improve the recognition accuracy of multiple breeds of dog breeds and achieves a brief application.The main research work is as follows:Firstly,based on the study of general image classification of deep learning,a dog breed recognition on the coarse-grained level is proposed.Combining the design of transfer learning technology with model fusion,using four commonly used convolutional neural network models to extract features from randomly selected partial pictures,and choose the two best performing models,Inception?v3 and Resnet152?v1 for dual model fusion,which use transfer learning to training on dog breed images.Then,the data set will be sent to the fusion network after beingidentified by the YOLO object detection algorithm to locate the target area,so as to improve the ability of object positioning.For 120 types of dog pictures,a training network model can achieve a learning accuracy of 93.02%.The classification accuracy on the test set can reach 73.2%.Secondly,in order to further improve the classification accuracy of dog breeds,this paper proposes a dog breed image classification based on multi-scale feature fusion on a fine-grained level.By using the feature pyramid structure to fuse shallow and deep feature maps,it generates strong semantic information on all scales to locate discriminative detailed targets.After that,select three area maps which have the best performance to determine the true category of the image,merging with the global features of the image to jointly predict the breed of dog.The key of this method lies in the end-to-end network architecture,which can automatically locate part level and obtain corresponding features.It reduces the need for high-cost manual labeling.After model training,the classification accuracy rate on the test set reached 83.5%.Experimental results show that the method can find more detailed information between different subclasses and achieve higher recognition accuracy.Finally,in order to verify the usability of the algorithm,a picture recognition audiobook based on the dog breed identification algorithm obtained in this paper was designed on the PC side,which completes a simple application for our dog breed identification.That is,by imputing an image,on the basis of accurately identifying the dog breed name in the picture,the corresponding breed introduction is retrieved by intelligently connecting to the corresponding database,and the output result is vocalized by voice synthesis to implement the science popularization reading of the corresponding breed,which realizes the intelligent automatic classification of dog breeds.It has far-reaching significance for animal science and even the protection and propaganda of endangered animals.