- 期刊
An Interactive Deep Learning Method For Fine-grained Image Classification
摘要
Fine-grained image classification refers to the classification of subcategories based on the basic categories already divided. Fine-grained image classification is a very challenging research task because of the data characteristics of small inter-class differences and large intra-class differences. Based on the analysis and research of existing fine-grained image classification algorithms, a novel fine-grained image classification method based on an interactive deep learning is proposed. First, YOLOv5 is used as the backbone network to improve the classification performance, and a random elimination enhancement selection strategy is designed. The feature elimination branch and feature enhancement branch interactions promote the network to learn more relevant information and capture potential distinguishable features. Then, a global diversified module is proposed to model the feature maps of different levels to improve the ability of network comparison cues. Finally, the internal standard imprinting data set is established, and the fine-grained algorithm is applied to the authenticity identification work to realize the practical application of fine-grained image classification in natural scenes. Model training can be efficiently trained in an end-to-end manner without bounding boxes and comments. Experimental results show that the accuracy of the proposed algorithm on three fine-grained image datasets, namely, CUB-200-2011, Standford Cars and FGVC-Aircraft, reaches 90.6%, 95.9% and 95.8%, respectively.