Research On Object Detection Method Of Remote Sensing Image Based On Deep Feature Fusion And Attention Mechanism

Compared with object detection for natural images,the technology for remote sensing images is different in data and application.For data,it mainly aims at remote sensing images of large scenes taken by lowaltitude satellites,and the shooting angle is from the top view.For application,it is mainly used in military object detection,ship search and rescue,map drawing and other fields.Therefore,research on object detection technology for remote sensing images is significant.Starting from the existing object detection technology,this article improves the technology by considering the characteristics of small targets and dense targets in remote sensing images.The main contributions are as follows.(1)Propose a dual feature pyramid network based on the attention mechanism.The traditional feature pyramid has shortcomings of rough up-sampling and different levels of abstraction between feature layers that cannot be directly added.This network uses spatial attention mechanism and channel attention mechanism to strengthen the object feature representation and improve the performance of object detection models.It contains a spatial attention module and a channel attention module.The former uses spatial attention mechanism to learn up-sampling position offset and sampling weights,and reconstructs shallow features from top to bottom,while the latter uses channel attention mechanism to reconstruct deep layers from bottom to top feature.Experiments show that this network can increase the detection accuracy by one percentage point.The experiments prove that the spatial attention module is helpful for small target recognition,and the channel attention module is helpful for large target recognition.(2)Propose a feature extraction network for regions of interests(Ro Is)based on feature fusion.The extraction of Ro Is is the key to object detection model.The traditional method takes Ro Is at different levels for same proposal but only selects one for subsequent process,which causes insufficient use of information.This network uses self-attention mechanism to re-learn the weights of the feature layer by fusing Ro Is at different levels,and adjusts the features using weights to obtain new Ro I.Because the final Ro I contains both detail information of shallow features and semantic information of deep features,the performance of object detection model can be improved.Experiments show that this network improves the detection accuracy by one percentage point.(3)Propose a weighted loss function based on Intersection over Union(Io U).Existing object detection models treat the positive samples whose Io U is greater than a certain threshold equally when calculating the loss,which causes loss is same when misclassification.However,positive samples that intuitively overlap more with the ground truth are misclassified is more unacceptable.Based on the point of view,this method designs a weighted loss function based on Io U.This loss function introduces Io U weighting factor in the classification loss function and the bounding box regression loss function,which is effective for guiding network training.Experiments show that the weighted loss function based on Io U improves the detection accuracy by two percentage points.The three technologies proposed in this paper improve different stages of the object detection model.First plays a role in the image feature extraction stage.Second plays a role in Ro I feature generation stage.Last is used for guiding Network training.Finally,the object detection model trained by combining these three detection technologies has significantly improved performance compared to Feature Pyramid Networks algorithm.