Research On Detection Algorithm For Specific Objects In High-resolution Optical Remote Sensing Images

In recent years,with the rapid development of remote sensing technology,the spatial,temporal and spectral resolution of remote sensing images is improving continuously,which makes more and more applications based on them impossible,such as environmental investigation,urban planning and regional monitoring and so on.Object detection in remote sensing images,as a basic task,has become a research hotspot.In particular,after rapid development,the high-resolution optical remote sensing images contain more detailed information,which makes object detection in this field more meaningful and challenging.But at present,the object detection methods in remote sensing images are mainly based on image processing and traditional machine learning,which both need rich experiences and complete prior knowledge.And most of them are just effective in specific environment,so they have poor scalability.On the other hand,the object detection in natural images has a fundamental improvement after the raise of the deep learning framework,nevertheless,deep learning is still in its infancy in the field of object detection in remote sensing images.In addition,current object detection methods in remote sensing images are based on the images cropped down to small sizes,but apparently,to detect object directly in the large-field ones has more realistic significance.Therefore,this paper proposed a size-scalable object detection algorithm in remote sensing images based on deep learning framework with scene feature fusion,and verified on different scales of high-resolution optical remote sensing images.Specifically,the algorithm proposed in this paper can be divided into three stages: for faster detection,firstly to get the possible locations of objects by a region proposals algorithm,and then extract features from each proposal,and classify them finally.For the stage of region proposals,by analyzing the particularity of object detection in remote sensing images,the size-scalable BING(Binarized Normed Gradients)algorithm was selected.Besides,this algorithm was improved by integrating multi-weak features scoring,and experiments results show that the improved algorithm achieved better detection rate and more accurate object coverage when obtaining the same number of candidate windows.For the stage of feature extraction,the deep learning was introduced into the object detection in remote sensing images.Concretely,we used convolutional neural network to extract deep features of candidate windows and the windows' context scene respectively,and then fused the two kinds of features for detection,which improved the detection performance.In addition,we solved the problem of the insufficient annotations on remote sensing images by transfer learning,which reduced the risk of over-fitting,and improved the network's ability of feature expression on remote sensing object and scene.Finally,for the stage of classification,we combined hard negative mining strategy when training the classifier,and the duplicate detection results was filtered to further optimize the results.