Research And Application Of Object Detection For Optical Remote Sensing Image Based On Deep Learning

With the rapid development of remote sensing technology,the application range of remote sensing images is becoming more and more extensive.In recent years,the resolution of optical remote sensing images has increased rapidly,and the amount of remote sensing image data has increased sharply,which makes it possible to detect fine objects in remote sensing image.The application of remote sensing image is no longer limited to image classification.However,the existing object detection technology for remote sensing image has not met the requirements of practical application in terms of accuracy,versatility and efficiency.At the same time,comp uter vision technology based on deep learning is developing rapidly.It has achieved much better results than traditional algorithms in many fields.It has many advantages,such as high accuracy,strong adaptability and so on.It has been used in many occasions,such as face recognition,pedestrian detection,intelligent monitoring and so on.Therefore,in order to improve the performance of the existing remote sensing image target detection technology,we study on object detection algorithm for optical remote sensing image based on depth learning in this thesis.Based on the region-based deep learning object detection algorithm,a high-precision and multi-class object detection method(MS-FRCNN)is proposed,which is suitable for high-resolution optical remote sensing images.Aiming at the problem that the detection model is easy to miss detection which the scale is small or scale varies in the remote sensing image,the idea of the feature pyramid network is used to detect in multi-layer features.Aiming at the challenges of complex background and deformation of object,the deformable convolution network is used to better extract the characteristics of the object itself and reduce background interference.In addition,the improved non-maximum suppression algorithm is used to deal with the overlapping bounding box.Experiments show that the improved method has the advantages of high accuracy and generalization ability,and performs well on the open optical remote sensing image data set.In addition,due to the interference from similarly shaped objects in remote sensing images,some kinds of targets are easily misjudged.Aiming at this problem,a target detection method MSS-FRCNN is proposed,which combines semantic segmentation information.Pixel-level classification information obtained by semantic segmentation while performing conventional target detection,and filtering unreasonable or undesired detection results according to objective correlation between target and environment,which effectively reduces the false detection rate.