Remote Sensing Images Object Detection Based On Improved YOLOv3

With the rapid development of digital image technology and computer vision technology,the object detection of remote sensing image has become the focus of current research.However,in the field of remote sensing image,there are still many challenges for rapid and accurate object detection.First of all,due to the special imaging perspective,the direction of target will be different,different directions will affect the accuracy of detection.Secondly,the remote sensing image has large pixels,but the small detection target is dozens or hundreds of pixels.In the process of object detection,it is difficult to extract the characteristics of the detected target.Finally,in the remote sensing image,some targets will be densely arranged,which will affect the accuracy of detection.Aiming at the problem that the traditional algorithm is difficult to detect accurately due to the dense array of objects in remote sensing images and the large difference in the size of objects,this paper proposes an improved target detection algorithm for remote sensing images based on the theory of deep learning algorithm.The main work contents are summarized as follows:1.In the aspect of data preprocessing,a variety of data enhancement methods are used to expand the data set and improve the clustering algorithm to increase the robustness and generalization of the algorithm for remote sensing images.The data set was expanded by means of contrast enhancement,brightness transformation,color balance and Angle rotation,so as to increase the robustness of the network model.Because of the characteristics of remote sensing image,the anchor box of remote sensing image is re-clustered by using the optimized K-means algorithm,so that the size of the anchor box is more consistent with the truth object and the positioning is more accurate.2.In the aspect of detecting frame positioning,gaussian model is introduced and NMS algorithm is improved to make the position information and size of bounding boxes more accurate.Using soft-NMS algorithm to replace traditional NMS algorithm,when selecting bounding boxes,it is no longer necessary to directly delete the bounding boxes with the highest confidence,which reduces the missed rate of densely arranged targets,improves the recall rate of the algorithm,and makes it have better detection results for densely arranged detection objects.Gaussian model is used to calculate the confidence of the coordinate of bounding box,so as to increase the accuracy of the bounding box.3.In terms of YOLOV3 algorithm,The loss function of YOLOv3 algorithm is improved.GIo U loss function is adopted and confidence loss of bounding box coordinates is introduced,and GIo U loss replaces MSE as bounding box regression loss function.GIo U can accurately reflect the degree of overlap and can directly minimize the distance between the two detection boxes.Meanwhile,it also takes into account the consistency of the aspect ratio between bounding boxes and trurh object boxes,which makes the detection and recognition of YOLOv3 algorithm more accurate.