Incremental Detection Methods For Object In Remote Sensing Images Driven By The Importance Of Sample

The intelligent interpretation and object detection of remote sensing images play a very important role in urban management and planning,activity identification,military and other application fields.Deep learning has been widely used in remote sensing image object detection because of its powerful multi-layer feature extraction capability.The classic deep learning paradigm requires training on datasets with complete features.With the development of remote sensing technology and the change of remote sensing application requirements,it is necessary to obtain more rich features of ground objects from more diverse remote sensing image data to adapt to complex and diverse objects detection scenes and tasks.Under the background of increasing remote sensing image data,the classical learning paradigm repeatedly learning all existing data consumes much computing resources and storage space.Remote sensing image object detection requires an incremental learning paradigm that can continue to learn new knowledge on the basis of existing ones,so as to integrate object knowledge efficiently and at low cost,and improve the versatility and generalization of the detection model.Therefore,from the perspective of remote sensing image objects detection and incremental learning,this paper analyzes the different effects of old and new samples in the scenario of instance / class increment on the model detection ability,and proposes remote sensing instance-incremental detection method and remote sensing class-incremental detection method driven by the importance of sample.The main work and innovations of this paper are as follows:(1)This paper studies the incremental learning mechanism of remote sensing instance-incremental task and remote sensing class-incremental task.According to the task requirements,precision evaluation indexes are set for remote sensing instance-incremental detection method,precision evaluation indexes and the model’s stability and plasticity performance evaluation indexes are set for remote sensing class-incremental detection method.The difference of learning value under remote sensing instanceincremental task and the conflict of learning knowledge about new class between the historical learning experience in the original detection model and the label guidance signal in the newly added data under remote sensing class-incremental task were explored respectively.The results of exploratory experiments are used as the theoretical support of research methods.(2)To solve the problem of low learning efficiency caused by the failure of learning model to recognize the difference of sample importance in the instance-increment detection task,a ranking-based instanceincremental learning method with sample importance for remote sensing detection model is proposed.Based on the uncertainty and inaccuracy of the predicted results,the importance of samples is measured and the rankscore is calculated.According to the rank-score,important samples in old instance data and new instance data is effectively learned from the data learning order and training contribution weight for model.The effectiveness of the proposed method is verified by experiments such as accuracy comparison and ablation analysis.(3)To solve the problem that the historical learning experience that identifies the objects of new class as background interferes with the normal learning of the objects of new class,a knowledge selective-distillationbased class-incremental detection method for remote sensing is proposed.At the input end of the knowledge distillation structure,a selective mask is added to the object of the new class,and when aligning the output of the teacher model and the student model,the classification result of the student model is reconstructed to calculate the selective distillation loss.By limiting the scope of knowledge distillation,the normal learning of the object of the new class is ensured,so that the remote sensing image object detection model has the ability to detect new and old remote sensing objects at the same time.The effectiveness of the proposed method is verified by experiments such as performance comparison and hyperparameter analysis.36 figures,26 tables,93 references...