Deep Active Learning For Fine-grained Visual Categorization

The sub-class recognition problem under the meta-class has the structural characteristics that the key features exist in the local area,and the solution to the finegrained visual categorization depends on the selection and fusion of key local information.Fine-grained recognition is a typical problem with large variance within classes and small variance between classes.The thesis focuses on the selection and fusion technology as the fulcrum,and designs the selection and fusion technical solutions at two levels: at the sample level,design sample-level semantic granularity and spatial location selection and fusion neural network representation;at the sample set level,select the batch-mode samples with a large amount of multi-dimensional feature are labeled online to solve the problem of the scarcity of fine-grained labeled samples.The specific work of the selection and fusion of two-level key features and samples is as follows:1)Sample-level stereo semantic selection and fusion neural network topology design: Propose a Cross-Granularity Multi-Attention Network(CGMA-Net),which applies multi-head attention operators to feature spaces at different abstract levels,and extract the fine-grained features of different receptive fields and different abstract levels,so as to express the local receptive fields of dynamic attention in the spatial domain,and express the concrete and abstract conceptual relationship in the conceptual domain;and then further integrate features at different granularity levels through the multi-head attention mechanism.The fine-grained information can realize the dynamic extraction of cross-granularity features.Under the condition that only category labels are needed,the model's feature extraction capability for fine-grained images and the integration capability between features at different granularity levels are enhanced.2)The design of learning mode for key sample selection and fusion at the sample set level: by decoupling the fine-grained recognition network and the unlabeled sample evaluation network based on deep active learning,relying on the rich semantic features generated in the recognition network,design a sample evaluation network with a simplified neural network topology.Efficiently realize the uncertainty measurement of unlabeled samples.At the same time,in order to adapt to the characteristics of batchmode data training in deep learning,the multi-feature aware deep active learning batch sampling module is designed,and samples which the model is highly confident in are given pseudo-labels in the semi-supervised learning.Combining two learning methods to alleviate the problem of insufficient annotation data in fine-grained visual categorization.3)Based on the above two theoretical results,the method proposed in this thesis has been experimented and compared on the public dataset CUB-200-2011,FGVCAircraft and Stanford Cars and the selection and fusion of key features experiments,which shows the advantages of the model's network design in recognition performance by.The realization based on deep active learning and semi-supervised learning methods and fine-grained image recognition framework proves its ability to improve the sample utilization of fine-grained image recognition problems.Finally the method proposed in this thesis are verified the feasibility and effectiveness.