Research On Few-Shot Object Detection For Remote Sensing Images

Few-shot remote sensing image object detection is a class of computer technology that has emerged in recent years,i.e.,a technology-oriented to real applications through the combination of computer vision object detection technology and few-shot remote sensing image applications.The current deep learning-based remote sensing image object detection algorithms have been developed for several years and have solved many practical problems.However,most deep learning-based algorithms need to be trained with a large number of visible data labels to achieve the expected detection results,and when the training is not sufficient,it is difficult for the training model to achieve the detection expectations.Therefore,few-shot object detection of remote sensing images is more challenging compared to traditional remote sensing image object detection.In this thesis,in order to improve the accuracy of few-shot remote sensing image object detection,based on the object detection,the corresponding solutions are proposed for the characteristics of remote sensing images and the method of running few-shot learning,the main work of this thesis are:(1)In this thesis,we propose few-shot object detection via context-aware aggregation for remote sensing images.Specifically,an improved algorithm of pixel attention-context-aware pixel aggregation is proposed for the problem of large-scale variation of objects in remote sensing images,which effectively alleviates the problem of detection errors due to large-scale variation of objects.A context-aware feature aggregation module is also proposed to increase the context awareness among features by performing graph convolution operations on each feature layer,which leads to better detection of smaller and dense objects.(2)In this thesis,we propose a few-shot object detection via contrast learning and adaptive optimization for remote sensing images.Specifically,for the situation that under-trained novel class samples are easily and incorrectly detected as fully trained base class categories in the case of small samples,this work proposes an adaptive data enhancement method to balance the training degree of base class samples and novel class samples by iterating hard-positive novel class samples.A region-specific comparison module is also introduced to use the contrast learning method to map the features between categories to a high-dimensional space during the training process,expanding the inter-class spacing and narrowing the intra-class spacing,so that the model can better judge the classification of categories and effectively alleviate the problem of error detection.This work also proposes a cascade attention mechanism,which combines coordinate attention and multi-scale pixel attention to solving the problem of partial image miss-detection and false detection by using background information and spatial information around the instance.In the few-shot setting,this thesis conducted experiments on public datasets by context-aware pixel aggregation to alleviate the detection error problem caused by large changes in object scale,enhance the robustness of the model,and improve the detection accuracy;on this basis,a graph convolutional neural network is introduced to enhance the feature context-aware feature information to further improve the detection accuracy.The adaptive data enhancement method strengthens the training of hard positive novel class samples,balances the training degree of base class samples and novel class samples,and improves the detection accuracy;on this basis,a region-specific comparison module is introduced to increase the inter-class gap of different classes and reduce the intra-class gap of the same class to further improve the detection accuracy;finally,a cascade attention mechanism is proposed for the image miss-misdetection problem,which makes full use of background information and instance The detection accuracy is further improved by making full use of the background information and the spatial information of the surroundings.The proposed method in this thesis has demonstrated its effectiveness by outperforming the current state-of-the-art few-shot remote sensing image object detection model in the widely-used m AP evaluation metric.