Remote Sensing Image Object Detection Based On Data Augmentation Strategy

In recent years,the theory of deep learning has developed rapidly and has made great progress in image classification,segmentation,detection and other tasks.Compared with the traditional object detection algorithm,the features extracted by deep learning are more complete and detection accuracy is higher.However,deep learning still has many problems to be solved.We found that in the algorithm of object detection,the detection accuracy of objects with large scale changes is still poor.In the image,the number of these objects are small and fit the long tail distribution.It’s why the result poor.Image data augmentation provides a powerful means for alleviating data overfitting of deep learning and improving the generalization ability of models.The existing data augmentation methods for object detection can be roughly divided into two types: No-boundingbox level and boundingbox level.No-boundingbox level is mainly for data pixel transformation(brightness,contrast,noise,etc.),while boundingbox level mainly transform the whole image,including clipping,multi-scale transformation,etc.However,these methods do not take into account the relative scale changes of the object in the image,and the newly generated image has no significant content changes compared with the original image.This thesis focuses on the remote sensing image object detection in deep learning,and using data augmentation to solve the problems existing in large-scale object detection.In this thesis,two kinds of augmentation methods are designed including image side data augumentation and latent feature side data augmentation.In order to evaluate the performance,we experiment on several datasets and get good performance.The main contributions of this thesis are as follows:1.This thesis proposed a data augmentation method by over-sampling object to augment the number of objects at different scales in the dataset.This method starts from the data source.At first,a predictor is set to dynamically predict the difficult objects in the dataset and add them into a object set.Then,according to the Alias method discrete distributed sampling algorithm,we generate a sequence matching the scale distribution of the original dataset and transform dataset size according to the sequence.Finally,a new composite image is synthesized according to the object sets.We conducted experiments on self-built Plane-22 datasets,Air-Sarship datasets,and CAESAR-Ship datasets using the Faster R-CNN algorithm.After using the oversampling amplification algorithm in this thesis,the performance in average precision is improved from 52.8% to 56.9%,27.5% to 29.3%,and64.2% to 65.1%.2.This thesis proposed a linear interpolated feature augmentation method for object detection task to find the potential object features of the same category in the feature space.Firstly,the input image is sent to backbone network to extract the feature map of the image,and then the area of the positive sample in the feature map is mapped into the feature vector,and then the linear interpolation method is used to make the disturbance between the features.Finally,the new composite feature and the original feature are concatenated into the branch of classification in object detection network.We conducted experiments on plane and vehicle classes of DIOR dataset using FCOS algorithm.After taking the linear interpolation feature augmentation method in this thesis,the performance in average precision was improved from 39.0% to 39.8% and 28.8% to 29.1%This thesis based on the deep learning method of remote sensing image object detection in the large scale changes and difficulty object proposes a object sampling image method by augmentating dataset and based on linear interpolation in feature space to augmentate data.It can effectively improve the performance of object detection in large scale change objects.It provides a new practice for remote sensing image object detection.