Research On Attention-Based Fine-Grained Image Recognition Method

Fine-grained image recognition is one of the most challenging tasks in the field of computer vision in recent years.It has important research significance and social value for the industrialization and practical application of machine vision methods,and can promote the intelligent development of various fields in social life.Compared with coarse-grained image recognition in the traditional sense,it has broader application prospects.Whether in the academic frontier fields such as medical imaging and biological research,or in the fields of industrial life such as urban transportation and manufacturing,fine-grained image recognition methods have important research values.Due to the huge difference between the fine-grained image itself and the traditional image,the use of a general convolutional neural network-based image recognition model to complete the fine-grained recognition task has great limitations and cannot meet the ideal target requirements.Therefore,combined with the specific characteristics of fine-grained images,the visual attention mechanism is introduced to effectively use the key information in the sample.On this basis,a fine-grained image recognition method based on aggregated attention is proposed.By combining channel attention and spatial attention mechanisms,it can effectively mine the detailed discriminative information contained in fine-grained targets.At the same time,reduce the amount of additional parameters introduced into the attention distribution,reduce the computational complexity of the recognition model,and seek a balance between the recognition effect and model performance.In order to improve the feature learning and expression capabilities of deep convolutional networks,a cross-channel global loss function is proposed to focus on the local discriminative features of the feature matrix to better guide the iterative training of the model.The experimental results show that compared with other finegrained image recognition methods,the fine-grained image recognition method based on aggregated attention proposed in the paper has better recognition effect on multiple data sets,which verifies the effectiveness of the proposed aggregated attention method.In addition,fine-grained image acquisition is generally difficult and needs to consume plenty of resources and time for data annotation.Thus,the number of data samples prepared for model training is relatively small which brings fine-grained image recognition tasks the problem of sample imbalance.In response to this problem the paper proposes a mixed method based on virtual sample enhancement for data expansion by generating virtual samples based on radiation transformation and random mixing methods,and adding them to the original data to assist model training.The robustness and generalization ability of the fine-grained image recognition model can be effectively enhanced.Through the calculation and comparison of the similarity within the classes and the differences between the classes of the generated virtual samples,the distribution consistency of the generated virtual samples and the original samples is verified.By comparing the recognition accuracy of the model on the same data set before and after adding virtual samples,the effectiveness of the virtual sample generation method used in the paper is verified.Furthermore,the virtual sample can also reduce the recognition bias of the model to a certain extent,which helps the fine-grained image recognition model to exert better performance in the face of unknown test samples.