Research On Fine-grained Recognition And Algorithm Based On Discriminative Triplets And B-CNN Model

As the saying goes:"workers want good things must first sharpen their tools." With the rapid development of Internet technology and computer technology,the image recognition and classification technology has also been developed rapidly.The general level of image classification can not meet the needs of people.The classification tasks are based on the basic concepts of animals,birds and flowers,cars,etc.Categories A simple rough classification develops into sub-categories of subcategories of this basic category.The difficulty of fine-grained image classification is that subclass differences are less obvious and subtle than base classes,but the classification of fine-grained objects can be very helpful to the production and life of today's society.Therefore,how to obtain satisfactory fine-grained classification accuracy is a highly concerned issue in the field of computer vision.The development of fine-grained image recognition and classification algorithms has generally formed two main classification methods.One is the fine-grained classification method based on component model and the other is the fine grained classification method based on deep convolutional neural network.Based on these two methods,this paper designs and implements a fine-grained image recognition and classification algorithm based on the discriminant triad model and a fine-grained image recognition and classification algorithm based on the B-CNN model.Fine-grained image recognition and classification algorithm based on discriminative triplets model is proposed in this thesis.Firstly,the nearest neighbor algorithm is used to construct multiple groups of rough aligned image sets,then the images are segmented in the form of sliding windows.The difference scores are mapped to the potential difference regions as discrimination patches,and the scores are selected based on the geometric constraints The highest six patches randomly construct candidate triples.At the same time,a new method based on the difference score map to propose a regional proposal can be very effective to locate the object to be identified.Finally,the maximum response of these selected candidate triples is concatenated to construct the middle-level image description of the whole feature of the target object and sent to the SVM classifier or GoogLeNet for training.Fine-grained image recognition and classification algorithm based on B-CNN model is proposed in this thesis.Due to the weak supervision training model can greatly reduce the workload of manual annotation,fine-grained image recognition and classification has always been the main direction of development.The fine-grained image recognition and classification algorithm based on B-CNN model uses only the labels of training samples,does not require any additional labeling information,optimizes and improves the traditional B-CNN network structure,and adds in the bilinear convolutional extraction process.The geometric constraints between components increase spatial feature information,further enriching the features extracted by the network.Based on the two algorithms presented in this thesis,we conducted fine-grained experiments on three datasets of 14-Class BMVC Vehicle Data Set,196 Stanford Automotive Data Sets and the University of California at Riverside Fine-Grained Data Set.The accuracy of the classification is up to 94.8,92.12%and 86.36%respectively,which verifies the effectiveness of the proposed algorithm.For some of the innovative ideas put forward in this paper,we also conducted relevant experimental verification,obtained some experience and further proposed the direction that can be further improved.