Object Detection In Remote Sensing Images Based On Deep Convolutional Neural Network

With the development of drones and the diversification of remote sensing observation platforms,more and more attention has been paid to the processing technology of remote sensing images.Among them,object detection technology based on remote sensing images has been widely used in military reconnaissance,terrain exploration and post-disaster reconstruction and became a research hotspot in this field.However,due to the inherent shortcomings of the remote sensing image imaging method,some objects in the image are small,angle variable and even some objects are in occlusion or complex backgrounds(the background and the object have similar texture and color),which brings great challenges to object detection.In recent years,although the deep learning-based methods has greatly improved the accuracy and speed of object detection,there are still some problems such as high false detection rate and missed detection rate for small object detection under complex backgrounds.Aiming at the above problems,this paper proposed a multi-component fusion network to improve the accuracy of object detection in remote sensing images.Firstly,a dual pyramid fusion network is constructed as a backbone network for feature extraction.This network is effective for low-level features.It can aviod the lack of spatial position information of the objects contained in the feature map of the convolutional neural network as the depth deepens,and improves the feature extraction capability of small objects in remote sensing images.Then the concept of relative intersection of union(IoU)is proposed and applied to the relative region proposal network,so that the network can maximize the local characteristics of objects according to the size of relative IoU,which is helpful for the detection of objects under occlusion.Finally,the contextual information network is used to learn the strongly correlated background features of the preliminary detection results from the previous network,which improves the accuracy of object detection under complex backgrounds.The experimental results show that the model performs well in detecting small objects in remote sensing images,even the objects are in occlusion and/or complex backgrounds.In order to further accelerate the speed of model detection,this paper also proposes a inception parallel attention convolutional neural network,which is mainly composed of parallel multi-scale attention modules,contextual attention modules and channel attention modules.First,the outputs of three parallel modules at multiple scales are fused to obtain rich multi-scale features,contextual features,and non-local correlation features,and then deformable convolutions are used instead of traditional convolutions in the resulting fused feature map,thereby more effective feature extraction for direction sensitive objects.Finally,the distance intersection of union(DIoU)loss is used to replace the traditional bounding box loss,which improves the model's convergence speed and obtains more accurate object position.The experimental results verify that using this network structure as the backbone for object detection can guarantee higher detection accuracy while effectively improving the detection speed compared with the previous model,and also has a better result for objects in complex scenes.