The Research Of Fine-grained Image Classification Based On Bilinear CNN

Fine-grained image classification is a research direction of great interest in the domain of image classification,which aims to distinguish different subclasses under the same category.Since fine-grained images have more similar appearance and features,thus making classification more challenging.At the same time,it is increasingly used in a wide range of fields,such as agriculture and medicine.so it is of practical importance to study fine-grained image classification.As more and more scholars have devoted themselves to the study of fine-grained image classification,many research approaches have occurred.The transition from traditional feature extraction approaches to feature extract ion using convolutional neural networks,in this paper,a bilinear convolutional neural network model is chosen as the object of improvement in the latter method to improve the classification accuracy,with the following main work:(1)To address the interference of irrelevant backgrounds in the dataset,the YOLOv4 algorithm is used to process the dataset,cropping out the irrelevant backgrounds in the detection results,leaving only the region where the target image is located,and using the cropped dataset as the training object for the subsequent classification model.(2)The idea of feature fusion is added to the bilinear convolutional neural network by adding two outer product operations to fuse the features extracted by the original outer product operation.Finally,it is sent to Softmax classifier for classification.The experimental results are compared with the original B-CNN,and the results are as follows: in the bird data set of CUB-200-2011,its classification accuracy has improved by 2.0%;On the Stanford Cars data set,it increased by 1.1%.(3)Changing a branch and classifier of a bilinear convolutional neural network,combination of VGG-M network with VGG-D network using DenseNet121 instead of the original bilinear structure,and SVM is used to replace Softmax in classifier.The data set processing is the same as the previous method.Compared with the original B-CNN,the experimental results are as follows: in the bird data set of CUB-200-2011,its classification accuracy improves by 1.5%;On the Stanford Cars data set,it increased by 1.2%.