Research On Object Detection Algorithms Of Optical Remote Sensing Images Based On Deep Learning

Optical remote sensing images play an important role in both military and civilian use,and can be used in disaster control,marine monitoring,urban planning,ecological environment monitoring,and many other fields.With the rapid development of remote sensing technology and the increase of satellites and spacecraft,we can obtain more optical remote sensing images with higher resolution.Small and medium-sized targets in the images are more clearly visible,which significantly increases the information content of optical remote sensing images and put forward higher requirements for the accuracy of its object detection method and technology.The object detection algorithm based on deep learning has achieved excellent results in the field of natural images target detection.However,because the imaging mechanism of optical remote sensing images is very different from natural images,the direct migration of target detection networks that perform well in natural scenes to optical remote sensing images object detection tasks often results in unsatisfactory detection.Therefore,this thesis mainly studies the optical remote sensing images object detection algorithm based on deep learning.Aiming at some of the key and difficult problems in the optical remote sensing images target detection,two advanced and applicable optical remote sensing images object detection algorithms are proposed,as follows:（1）Aiming at the problem of poor detection effect on small targets and densely arranged targets when the traditional Retina Net network is applied in the optical remote sensing images target detection task,an improved algorithm based on Retina Net is proposed.First,the feature extraction network is improved by using the idea of grouped convolution to enhance the feature extraction ability without increasing the amount of parameters.Secondly,the balanced feature pyramid module is added to the feature pyramid network FPN,the features of each layer of the feature pyramid network output are merged and enhanced,and the non-local neural network is used to refine the enhanced features to enhance useful feature information.Finally,GIOU loss is used to calculate the regression loss of the bounding box.The experimental results show that compared with the traditional Retina Net network,the improved algorithm of this thesis has greatly improved the detection accuracy.（2）Aiming at the performance of the anchor-based detection algorithm heavily dependent on the manually set a priori anchor parameters,which limits the generalization and robustness of the detection network,this thesis introduces the FCOS anchor-free algorithm and improves on it.By using the neural network structure search method to reconstruct the feature pyramid structure to generate better multi-scale features.Then the angle regression branch is added to the detection head of the FCOS network to enable it to detect the rotating object.The experimental results show that the detection accuracy of the improved anchor-free detection algorithm in this paper is superior to other detection algorithms.