Research On Optimization And Application Of Bilinear Convolution Neural Network

Fine-grained image recognition has a wide range of research needs and application scenarios in both industry and academia.Bilinear Convolutional Neural Networks(B-CNN)has excellent ability to extract fine-grained features due to its coding method.One of the methods to improve the accuracy of fine-grained image recognition.This paper focuses on optimizing the B-CNN model in two directions: first,this paper proposes a B-CNN model based on the Histogram of Oriented Gradient(HOG)feature extraction,which uses HOG to extract fine-grained features to strengthen the supervisory constraints on the network,and improves the activation function of the network model to make the model converge faster.In the second direction,this paper proposes a fine-grained image recognition based on multibranching and multi-scale,which adds localization branches and multi-scale branches to the basic network to enable the model to locate different distinguishing local features more accurately.The specific work of this paper is as follows:1.In this paper,we propose a B-CNN model based on HOG feature extraction to address the problems of improperly extracted features and improperly constructed activation functions of the network model.In this paper,the HOG feature map of the original image is extracted first,and then it is input to the B-CNN as a complement of image features for training.Meanwhile,the Swish-Arc Re LU activation function is constructed by combining Swish function and Arc Re LU function to further improve the network model.In this paper,this model is applied to CaltechUCSD birds(CUB＿200＿2011)dataset and Stanford-Dogs dataset,and the experimental results show that the recognition accuracy of this method reaches 91.3% and 97.9%,respectively,which is significantly better than the traditional B-CNN model and has practical application value.2.In this paper,we propose a multi-branch and multi-scale B-CNN model based on the problem of how to locate the distinguishing regions more accurately and how to enhance the data more effectively.In this paper,we use three branch networks to supervise the image recognition process,firstly,we extract the overall features of the object through the original branch,then we obtain the bounding box information of the object through the localization branch to localize the recognition target,then we extract several regions with the strongest differentiation and the smallest redundancy through the multi-scale branch to crop the multi-scale local image for data enhancement,and finally,we send the cropped image to the convolutional neural network model for training,so that the network model can learn fine-grained image features of different scales and different parts.The experimental results show that the multi-branch and multi-scale B-CNN model proposed in this paper is of practical application.