Object Detection Algorithm For Optical Remote Sensing Image Based On Deformable Convolution And Multi-model Fusion

Object detection is an important branch of research in the field of remote sensing,which has important applications in urban development and military affairs.With the continuous development of deep learning in recent years,deep learning-based methods have gradually replaced traditional object detection methods as the mainstream method,and have achieved extensive research results.In view of the characteristics of wide range of scale distribution of optical remote sensing images,uneven distribution,many small objects,complex background,and the inaccurate characteristics of current object detection algorithms.Based on the object detection method based on deep learning,this paper proposes an object detection method for optical remote sensing images based on deformable convolution and multi-model fusion is presented.The main work is as follows:1.In view of the characteristics of wide range and uneven distribution of object scales in remote sensing images,this paper proposes an improved SSD optical remote sensing image object detection method.Firstly,the fixed anchor box of the object detection algorithm is difficult to match the object with uneven distribution due to wide object scale distribution and uneven object distribution.This paper proposes an anchor box adjustment algorithm,which dynamically adapts the position of the anchor box to adapt to different object distributions.Secondly,this paper proposes a feature fusion module to fuse features of different layers to enhance the feature extraction capability of the network.Then,to solve the problem of many virtual frames in the traditional non-maximum suppression algorithm,an improved non-maximum suppression algorithm is proposed to further suppress the virtual frames and improve the accuracy.2.For the problem of low detection accuracy caused by dense objects in remote sensing images,we further study that this is due to the object detection algorithm cannot accurately extract the feature of the object area.In response to this problem,we propose to use deformable convolution to extract more Accurate object feature,and creatively proposed two-stage deformable convolution sampling point adjustment algorithm to extract more accurate features.In addition,we also proposed an attention module to enhance the features of the object area.Suppress the effects of background.3.There is a limit to the detection accuracy of a single model.When it reaches a certain accuracy,it becomes extremely difficult to improve it.This paper achieves a higher detection accuracy by using multiple detection models than using a single model.However,training multiple models independently requires a lot of time and greatly slows down the detection speed.Based on the Adaboost algorithm,this paper innovatively proposes a boosting multihead network(BMHN),In the case of sharing the same backbone among different models,maintain the accuracy of each detection model and maximize the difference in their detection results.Although the detection accuracy of each model is lower than that of the separately trained model,fusion of these multiple models can obtain higher detection accuracy and maintain detection speed.