The Research Of Fine-grained Image Classification Method Based On Bilinear Convolutional Neural Network

Compared with coarse-grained image classification,fine-grained image classification is based on the classification of fine sub-categories under large image categories.It has a wide range of applications in the fields of identifying endangered species and creating species recognition systems.It has become a computer vision in recent years.Research topics of interest in the field.For fine-grained image classification tasks,the difference between the classes of the images is large and the differences within the classes are small.The effect of using traditional convolutional neural networks for this classification task is general,mainly because the high appearance similarity between the images increases the classification The rate of misjudgment and the difficulty of classification.Bilinear Convolutional Neural Network(B-CNN)uses two-way convolutional neural network without the help of part labeling to coordinate the extraction of local features and information classification of objects,thereby improving the accuracy of finegrained image classification.The low-rank bilinear volume integral model(LRBP)based on B-CNN design can greatly reduce the dimension of the parameters,and it has become a model with better classification effect in the bilinear network model.However,the LRBP model does not effectively solve the problem of accurately locating key image regions with typical resolution under fine-grained image classification tasks,and the problem of effectively extracting distinguishable features from the detected key image regions for classification.The thesis studies the above problems based on the bilinear convolutional neural network LRBP model.The specific work of the thesis is as follows:(1)A fine-grained image classification model(LBCNN-RA)based on LRBP combined with residual and attention mechanism is designed.The LBCNN-RA model uses a deep residual network to solve the problem of gradient disappearance when the number of layers of convolutional neural networks increases.At the same time,it combines the deep residual network and the attention mechanism into a residual attention module,where the attention mechanism is part of the channel attention It is composed in series with the spatial attention dual module to obtain the attention information of the image in two dimensions of channel and space,thereby effectively positioning the key image area with typical resolution,and finally performing bilinear low-rank aggregation and using softmax for classification.The experimental results show that the classification accuracy of the LBCNN-RA model on the CUB-200-2011 data set is increased by 2.2% compared with the LRBP model,and the classification accuracy of the LBCNN-RA model is increased by 1.3% compared with the LRBP model on the Stanford Cars.(2)Design a fine-grained image classification model(RALBCNN-DS)based on LBCNN-RA that introduces two-scale fusion.The RALBCNN-DS model first uses the deep residuals in the LBCNN-RA model and the attention combined with the network for training,and then conducts bilinear network low-rank aggregation,and finally introduces the center loss metric learning method during network classification.The classification features of the two scales and the center loss are fused and learned,so as to effectively extract the distinguishable features of the detected key image regions from the two-scale dimension,and achieve better classification results.Experimental results show that compared with the basic model LBCNN-RA,the RALBCNN-DS model has a 0.5% improvement in classification accuracy on the CUB-200-2011 data set,and an improvement in classification accuracy on Stanford Cars by 0.7%,and finally compared with other current models.The classification effect of the fine-grained image classification model is compared.