Object Detection For Remote Sensing Images Based On Center Distance And Multi-scale Feature Fusion

Remote sensing images play an extremely important role in many fields such as agriculture,military,outer space exploration,and environmental monitoring.As one of the core branches of remote sensing image processing,object detection for remote sensing images has received much attention from many researchers and has a very broad development prospect.At present,the object detection algorithm based on convolutional neural network is the most mainstream method in the field of object detection.However,due to the many differences between remote sensing images and ordinary natural images,object detection for remote sensing images still faces many challenges.This thesis focus on remote sensing image object detection and propose an improved algorithm for small-scale object detection,multi-scale detection and local false anchors in remote sensing image object detection,which achieves better detection results.The main work is as follows.(1)To solve the problem that area-based IoU cannot accurately reflect the overlap between anchors in remote sensing images,a remote sensing image object detection method based on center distance and redundant anchors modifying NMS is proposed.The method combines the center distance and intersection over union to measure the overlap between the anchors,and uses both metrics to decrease the scores of the anchors.Finally,the information of the removed redundant anchors is also used to modify the retained anchors to improve the accuracy.Experiments on different datasets have shown the effectiveness of the method.(2)To solve the problem of small object detection in remote sensing image object detection,a remote sensing small object detection algorithm based on deep feature elimination and channel attention is proposed.Firstly,the method makes the shallow feature map minus the deep feature map in the FPN to eliminate the deep features that do not contain small objects to reduce interference and make the small object features more significant.Then different levels of feature elimination for the other layers of the FPN are performed and their layer differences from the shallow layers are used as weights to attenuate the features of other scale objects in the shallow feature map.Finally,a channel attention mechanism is introduced to compensate for the semantic information loss in the shallow feature maps.The experimental results have shown the effectiveness of the method for remote sensing small object detection.(3)To solve the problem of multi-scale object detection in remote sensing image object detection,a remote sensing multi-scale object detection method based on multi-scale fusion and A-IoU loss function is proposed.The feature maps of each level output from FPN and filters the features of different levels through the channel attention mechanism,which satisfies the demand for different levels of information for different scale detection,are fully fused.An A-IoU loss function is also proposed to solve the problem of local false anchors which are commonly found in remote sensing multi-scale object detection.The A-IoU loss function enables the model to focus more on local false anchors during the training process to reduce their generation and thus improve the detection accuracy.Comparative experiments have shown that the method achieves good detection accuracy on most scale categories.