Object Detection In Remote Sensing Imagery Based On Deep Multi-scale Feature Learning

In recent years,the field of remote sensing has entered a new period of intelligent development.The field of remote sensing is closely integrated with artificial intelligence technology and has gradually gone deep in the military,civil,and commercial fields.Its development level has become one of the important signs to measure comprehensive national power.As an important research direction in remote sensing image processing,object detection in remote sensing imagery aims to classify and locate objects in remote sensing imagery.For the characteristics of multi-scale and multi-direction objects in complex remote sensing imagery,three object detectors based on the multi-scale feature learning theory in deep learning are proposed.The contents of this paper are as follows:(1)A method of object detection in remote sensing imagery based on semantic-guided feature enhancement and global context learning is proposed.Based on Retina Net detector in the baseline method,the method fully considers that there is a lack of semantic features of the multi-scale object in remote sensing imagery,as well as the background in remote sensing imagery complex problems.On the one hand,we build a semantic-guided feature enhancement module in remote sensing imagery to improve the expression of semantic features of multi-scale objects.On the other hand,we build a global context learning module for establishing context information association in remote sensing imagery.The two modules are integrated to fully learn the features and improve the representation ability of the features.The proposed method was experimentally verified on two sets of data sets NWPU VHR-10-v2 and DIOR.The experimental results showed that the mean average precision of the proposed method was higher compared with other methods.Compared with the baseline method,the mean average precision of the proposed method on the two data sets was improved by 4.01% and 2.94% respectively,which proved the effectiveness of the proposed method.(2)A fully convolutional object detection method in remote sensing imagery based on feature location alignment learning is proposed.To detect multi-scale objects more flexibly,the baseline method adopts the Fully Convolution Neural Networks object detection method in remote sensing imagery.On the one hand,in the case that the baseline method predicts the misalignment of the object location feature and the classification feature,the feature alignment module is introduced to make the classification feature of the object and the updated location feature of the object keep consistent,and alleviate the mismatch between the classification score of the object and the object position.On the other hand,to solve the problem of unstable optimization of position regression task,a kind of intersection ratio loss function of distance normalization is introduced to make regression training more stable and efficient.Through verification of the data set NWPU VHR-10-v2 and DIOR,compared with the baseline method,the mean average precision of object detection was significantly improved by 2.8% and 2.1% in the two data sets,respectively,proving the effectiveness of the proposed method.(3)A multi-scale feature enhancement and orientated bounding box object detection method in remote sensing imagery is proposed.To detect the target more accurately,the baseline method based on the rotated bounding box is used to detect the object.On the one hand,in the face of multi-scale objects in remote sensing imagery,the multi-scale feature enhancement module is introduced to improve the expression ability of features.On the other hand,objects in remote sensing imagery have the problem of low recall rate,so a better matching strategy between the anchor and real Ground Truth Box is proposed to increase the recall rate.The experimental result of validation on data set DOTA shows that the mean average precision of the proposed method is higher than other methods,and the mean average precision of the proposed method is 1.188% higher than that of the baseline method.