Design And Implementation Of Remote Sensing Target Detection Algorithm Based On Semi-supervised Learning

With the continuous development of remote sensing satellite technology,massive amounts of remote sensing image data are generated every day for environmental monitoring,ocean research,climate monitoring,resource exploration,etc.The task of target detection for remote sensing image is a hot topic in the field of remote sensing image processing and analysis,and it is also an important basis for the follow-up complex research tasks.In the task of target detection,data labels need to accurately mark the location of the target,and it often costs a lot of manpower and financial resources to complete all data labeling.Therefore,the introduction of semi supervised learning to effectively use unlabeled data to help model training,alleviate human dependence,and avoid data waste,has high practical significance for remote sensing target detection research.This thesis is divided into two parts:The first part is the network optimization design for the characteristics of remote sensing target.Due to the particularity of the shooting height of remote sensing image,remote sensing targets have the characteristics of multi-scale,multi direction,small and dense.According to its characteristics,this thesis optimizes the design of Faster R-CNN network: adding multi-scale prediction structure;using K-Means to generate anchor settings that conform to the dataset’s distribution;proposing spatial attention module SAM and channel attention module CAM,which are applied to feature maps of different levels in different order;modifying the regression box loss to GIOU loss that is more consistent with the evaluation index.After verification,the improved RF R-CNN network has a 3.84% increase in m AP compared with the original network.The second part is the research of remote sensing target detection algorithm based on semi-supervised.First,this thesis conducts a progressive research on the semisupervised self-training method.Firstly,a simple self-training experiment based on pseudo tag is carried out on the target detection task,and a slow start self-training method is proposed.Secondly,through the analysis of the limitations of slow start method,an active semi-supervised self-training method combined with active learning is introduced.In this process,this thesis proposes a committee-based uncertainty sampling strategy to sample samples with high uncertainty and low uncertainty for active semi-supervised self-training.The experimental results prove that it is very helpful to improve the performance of the model,and the m AP is 3.37% higher than that of the random sampling strategy.Second,research is conducted on a consistent regularization method that does not need to generate pseudo-labels.This thesis proposes a student-teacher semi-supervised training framework based on Mean Teacher.Under this framework,this thesis designs a consistency loss for target detection tasks,and simultaneously trains both labeled samples and unlabeled samples.Experimental results prove that when the method uses the same number of unlabeled samples,m AP is 2.11%higher than the improved slow-start self-training method.