Research On Few-Shot Fine-Grained Target Recognition Method Of Remote Sensing Images

With the continuous development and progress of earth observation technology,the resolution of remote sensing images has been increasing,which provides the data basis for remote sensing fine-grained target recognition.Remote sensing fine-grained target recognition refers to the ability to accurately distinguish different and similar subcategories within the broad category to which the target belongs.Fine-grained target recognition has important application value in fine agriculture,national defense and other fields.In the existing fine-grained target recognition methods of remote sensing images,fine-grained samples are mainly annotated manually,which is time-consuming and laborintensive,and the quality is more influenced by human subjective factors,resulting in limited data volume of fine-grained annotated samples of remote sensing images and facing the problem of few-shot.Although the few-shot method of remote sensing image fine-grained target recognition has made some development,but in the face of the complex background of remote sensing targets and diverse testing areas,there are still the following shortcomings:(1)the existing few-shot learning method of remote sensing image fine-grained target recognition does not fully consider the complex background of remote sensing data,making it more difficult for the method to locate the effective finegrained feature area,and the feature target(2)the existing few-shot learning methods for fine-grained target recognition of remote sensing images do not fully consider the diverse test areas of remote sensing data,and cannot guarantee that the training area and test area are distributed together,so the fine-grained features of the learned feature targets are not accurate in the test area.To address the above problems,this thesis proposes a training task sampling strategy that takes into account the fine-grained features of remote sensing and a dynamic remote sensing few-shot fine-grained recognition method with coupled meta-learning,which expands the original sample information by introducing the information of target categories so that each batch of training data contains only fine-grained samples belonging to the same category,converts the training task distribution into fine-grained task distribution,and obtains more accurate fine-grained recognition capability.The meta-learning training algorithm learns multi-region recognition prior and dynamically adjusts the fixed metric space to adapt to the distribution of new task sampling regions to improve fine-grained target recognition capability.The main points of this paper are as follows:(1)proposes a training task sampling strategy that takes into account the fine-grained features of remote sensing,and optimizes the distribution of fine-grained targets learned by the few-shot recognition method of remote sensing images.This thesis introduces the category information of fine-grained target broad categories,gives each sample a coarse broad category and a fine-grained category label,and converts the target distribution learned by the few-shot method into a fine-grained target distribution.By mining the small differences of fine-grained categories in each large category in the dataset,the fine-grained features of their corresponding large categories are extracted more precisely.To ensure that the tasks sampled in the few-shot fine-grained recognition method for remote sensing images belong to similar distributions,the task sampling strategy further improves the accuracy of fine-grained features learned by the few-shot granularity recognition method for remote sensing images by introducing a category generation method based on Beta distribution to ensure that each task has the same number of fine-grained categories;(2)proposes a coupled meta-learning dynamic remote sensing few-shot fine-grained recognition method that adapts to diverse remote sensing image data test regions.This thesis introduces a meta-learning training algorithm capable of learning multiple region recognition priors to dynamically learn prior knowledge on new regions over multiple regions.The proposed method divides the data in the task into two parts,one for computing the metric space and one for dynamically adjusting the space,and uses the meta-learning training algorithm to synthesize the prior knowledge learned from multiple task regions and adjust the distribution of new regions to achieve cross-region capability and significantly improve the accuracy of the fine-grained recognition method for few-shot of remote sensing images;(3)The proposed method is validated on a publicly available dataset and compared with several existing few-shot fine-grained recognition methods.The experimental results show that the proposed training task sampling strategy that takes into account the fine-grained features of remote sensing can significantly improve the extraction capability of fine-grained features of remote sensing images and realize the improvement of recognition accuracy; the proposed coupled meta-learning dynamic remote sensing few-shot fine-grained recognition method makes a significant improvement in accuracy with very few training samples.This shows that our method can be well applied to the few-shot problem in the target recognition task of remote sensing images,which helps to develop more high-precision remote sensing applications.