Object Detection And Direction Estimation In Remote Sensing Images Based On Deeping Learning

With the development of remote sensing technology,the available remote sensing images have become more abundant,and the object detection in optical remote sensing images has played an increasingly important role in the military and civilian fields.At present,although many advanced object detection algorithms have succeeded in natural images,there are two factors that limit their development in remote sensing object detection tasks: On the one hand,the detection accuracy is not high due to the large changes in the scale,rotation direction and distribution density of the object in the remote sensing image.On the other hand,the original remote sensing image is usually a high-resolution large-scale image,which causes unnecessary time consumption due to the detection of blank images during block detection.Therefore,further research is needed to enable efficient remote sensing object detection using rich high-resolution optical remote sensing image resources.Based on a thorough study of traditional remote sensing image object detection methods and deep learning related theories,this thesis focuses on the main difficulties of remote sensing image object detection and the problems of existing methods.taking typical remote sensing ship as an example,the method of object detection and direction estimation of large-scale remote sensing images based on deep learning is emphasized.This thesis proceeds from the following three aspects:(1)Analyzed the traditional remote sensing image object detection algorithm,and researched the remote sensing ship object detection and direction estimation based on local salient features.After analyzing its limitations,it led to deep learning.Further expand to introduce the basic theory of Convolutional Neural Network in deep learning,analyze the current popular convolutional neural network model structure and its advantages and characteristics.Then,we studied several typical object detection algorithms based on deep learning.After analyzing the detection principles and advantages and disadvantages of each detector,the object detector with excellent detection performance——Light-head R-CNN was selected as the research basis based on the needs of its own experiments.(2)A fast object detection method based on a coarse-to-fine strategy for large-scale remote sensing images is proposed.In the coarse detection stage,a classification model trained by transfer learning is used to obtain candidate region images from large remote sensing images.In the fine detection stage,an improved object detector is proposed for candidate region image object detection and direction estimation.The detector achieves more accurate object detection by using feature fusion,Inception module and deformable PSROI pooling.Aiming at the object direction estimation problem,the angle parameter is added while the detector predicts the coordinate parameters of the object position.By modifying the loss function and training,the detection model can achieve stable rotation direction estimation,so that rotating object detection can be achieved by rotating the bounding box.Compared with the traditional horizontal bounding box,the rotating bounding box is more compact,and it is more robust to background interference.(3)Constructed the classification data set and object detection data set needed for the experiment by itself.Three remote sensing object detection algorithms that can perform rotating object detection are compared in the experiment: improved SSD(DRbox),improved Faster R-CNN method,and R2 CNN ++.The results prove that the remote sensing object detection detector proposed in this thesis has the highest detection accuracy.Finally,through large-scale remote sensing image test experiments,it is proved that the large-scale remote sensing image object detection based on the coarse-to-fine strategy can effectively improve detection efficiency and achieve a balance between accuracy and speed.