Research On Microscale Object Detection In Optical Remote Sensing Image

Object detection in remote sensing image has important value in both military and civil affairs.At present,the performance of microscale object detection,such as vehicles and small boats in remote sensing image,is not satisfactory compared to large ground target.Due to the limitation of detection methods and technologies,the spatial resolution of optical remote sensing image is limited,and the microscale objects in the image are small in pixels.In the meanwhile,the features of the same class object vary a lot.Single image super resolution(SISR)and multi-scale feature fusion techniques are used in this paper to solve the above problems.By this way,the image resolution can be improved while the feature extraction ability of the neural network can be enhanced,and therefore can achieve improved performance of microscale object detection.The resolution of microscale objects in optical remote sensing images is low,and their features such as edges and textures are rare,which limit the accuracy of detection and recognition.In order to address this issue,the SISR based preprocessing method is proposed to enhance the image beyond its native resolution and restore microscale target features,so as to improve target detection accuracy.In addition,the super-resolved images of satellite imagery are often too smooth,and the high-frequency details of microscale objects are absent,while these high-frequency details such as edges and textures are vital for detectors.To handle this problem,perceptual similarity loss is proposed.By using it,high-frequency details in the images can be effectively generated,and that improves the accuracy of detectors.The feature extraction process of the deep convolutional neural network down samples the feature map continuously,the microscale object generally disappears in the deep feature maps which leads to the object detection performance drops sharply.So multiscale feature fusion is proposed to fuse deep and shallow feature maps to improve the performance of microscale object detector,and the anchor-free based method is also used to effectively reduce computational costs.This paper also proposes a novel end-to-end framework to train SR and detection task in the same time.In this framework,the object detection task can guide the front SR sub-network so that the super-resolved images are more suitable for machine vision which makes a better performance of microscale object detector.Finally,the contrast experiments are conducted made on public optical satellite imagery.By comparing the models proposed in this paper with other state-of-the-art detectors,the experimental results show that the proposed method has higher accuracy for microscale targets detection.