Research On Fine-grained Image Classification Based On Weak Label Data

Fine-grained image classification is one of the most important researches in the field of computer vision.Recently,with the development of deep learning technology,the effect of fine-grained image classification has been improved significantly.Because it still requires a lot of manually labeled image data that building these classification models,how to use a small number of samples or data accessible easily instead of labeled data is the focus of researchers.At the same time,a large number of weak label data can be easily obtained through the Internet platform,but it is hardly to use effectively because of the weak labels.In this respect,this paper proposes some methods including generate labeled data and training fine-grained image classification model using weak label data from the following aspects:1.Due to the inaccuracy of image data label which is obtained from the Internet,this paper proposes a weak label data labeling method based on confusion probability evaluation.After acquiring a large number of image data with text information on various search engines by using web crawler technology,the confusion probability of different types of samples is predicted by predicting classification model,and a large number of samples with correct label which are difficult to recognize by classification model are judged.Compared with the method of judging sample label directly,this method effectively reduces the waste of data resources and the cost of data acquisition.2.The traditional fine-grained image classification model requires that all kinds of image data are labeled in advance.To solve this problem,this paper only uses a small number of labeled data sets to initialize the classification model,and continues to optimize with a large number of weak labeled data from the Internet to achieve alternate data generation and model optimization,eliminating redundant time.3.The main difference between fine-grained images is the details of the categories,but the data sets do not contain the labels of the details usually.To solve this problem,this paper only uses the sample class label to detect the different parts of the image,which improves the accuracy of fine classification without additional information.4.In this paper,the image classification data set named "Chinese Food" is generated by using the web crawler and the data annotation method proposed in this paper.The data set which contains 20 kinds of food and totaling 73125 images can be used as standard dataset for fine-grained image classification.