Object Detection Technology Research In Visible Spectral Remote Sensing Images Based On Deep Learning

The vertical view of remote sensing images bring great convenience for people to observe objects on earth.As the advancement of aerial and spacial carrier technology,there’s huge demand on remote sensing data processing.As a typical image processing task,object detection is very important in the processing of remote sensing images.It is a hotspot that how to accuratly detect objects of interests.This paper’s research develops around object detection techniques of visible light remote sensing images based on deep learning.The main research results are as follows:1.Research data augmentation techniques targeting objects on remote sensing images.Designed a innovative data augmentation technique named Similar Tragets Replacing（STR）.First randomly select targets on training set images,then replace them into targets from similar categories,finally do affine trasformations on those similar targets.Experiments on DOTA dataset shows,STR can effectively improve training and detecting performance of deep learning algorithms.2.Study the angle prediction problem in rotate object detection,raised a method called distance sensitive densely coded labels（DSDCL）.For the problem of accuracy loss of the final prediction box due to the coded high-bit misdetection in the method of angle prediction using the coded classification method,the rotated anchor boxes are used instead of the horizontal anchor boxes.Angle difference between the prediction boxes and the anchor boxes is used as the reference for angle classification training,so that the number of samples with misdetection is reduced by minimizing the occurrence of 1 in the coded high-bit for positive samples with the help of the angle threshold of the positive sample anchor boxes.Experiments on the DOTA and UCAS-AOD datasets validate the effectiveness of the method;For the high-bit misdetection problem of the coding classification method,Bit Weight Enhencement Weights（BWEW）are designed to make positive augmentation to loss function in high bits.By doing so,prediction accuracy in high bits will increase and predicted boxes’ accuracy loss will be reduced.Experimentally,BWEW are proved that can significantly improve performance of angle classification algorithms;The difference between the angle prediction and the ground truth value and their encoded Hamming distance are not linearly related.The Coding Distance Compensation Weight（CDCW）is designed to pass the angle difference linearly into the angle classification loss function,so that the network learns the angle classification features quickly and accurately.Experiments demonstrate that CDCW can improve the accuracy of the angle classification algorithm.3.The algorithm of remote sensing object detection based on group equivariant convolution is studied.The group equivariant convolutional Faster R-CNN（GEC FRCNN）is proposed.Firstly,equivariant features of objects are obtained by using group equivariant convolution,and after feature recombination,the designed orientation-sensitive region proposal network is used as the RPN.The orientation-sensitive features of the objects are obtained by orientational object feature cropping.The experimental results show that the object detection algorithm based on group equivariant convolution has better detection result compared with the algorithm based on conventional convolution.