Weakly Supervised Object Detection Based On Multiple Instance Learning In Remote Sensing Images

With the development of deep learning and remote sensing technology,fully supervised object detection methods have made significant progress in the remote sensing field.However,these algorithms require a large number of images with instance-level annotations,which is very difficult to manually label remote sensing images.Compared to fully supervised object detection methods,weakly supervised object detection methods only require image-level annotation data for training,which has become a focus for many researchers.Many weakly supervised object detection methods currently using category activation maps or multiple instance learning have progressed in natural scene domains.Among them,weakly supervised object detection methods based on multiple instance learning can detect more complete objects and are the mainstream research direction.However,compared with natural scene images,remote sensing images have other challenges,such as large differences in object scales,difficulties in object localization,and arbitrary object angles.Therefore,weakly supervised learning methods for natural scene images cannot be directly applied to object detection in remote sensing images.To solve these problems,this thesis focuses on weakly supervised object detection based on multiple instance learning in remote sensing images,and conducts research from three aspects:feature fusion,curriculum learning,and multi-angle selective search.The main research work and results are summarized as follows:(1)Current feature fusion methods use the same way to fuse features for all objects,which can extract the object features of different scales,but ignores the fact that different objects pay attention to different features,resulting in fused features that cannot effectively represent objects.To address this issue,this thesis proposes a weakly supervised object detection method based on dynamic feature fusion in remote sensing images,which designs a Dynamic Feature Fusion Network(DFFNet).When fusing high-level and low-level features,the model dynamically adjusts the fusion convolution kernel using image features to enhance its ability to dynamically extract different scale object features,thereby improving its detection performance.The experimental results show that the average accuracy(mAP)metrics of DFFNet are respectively improved by 1.04%and 2.44%on DIOR and DOTA datasets,compared to the state-of-the-art methods.(2)Weakly supervised object detection methods currently use remote sensing images with different levels of object localization difficulty for training,which reduces the detector’s performance.To address this issue,this thesis proposes a weakly supervised object detection method based on adaptive curriculum learning in remote sensing images by using multi-stage training and curriculum learning based on DFFNet.This method integrates the differences in object scales and localization confidence to divide the remote sensing images into different courses.The model adaptively learns from simple to difficult data to gradually improve the model’s detection ability through curriculum learning.The experimental results show that the model’s detection performance continuously improves with the advancement of adaptive curriculum learning.The mAP metrics are respectively improved by 5.62%and 3.16%compared to DFFNet,and by 6.39%and 6.09%compared to the state-of-the-art methods on the DIOR and DOTA datasets.(3)Aiming at the problem that weakly supervised object detection models cannot effectively extract features of rotated objects and accurately locate objects at different angles,this thesis proposes a weakly supervised object detection method based on multi-angle selective search on the basis of the DFFNet model in remote sensing images.This method uses the multi-angle selective search algorithm to generate candidate boxes that conform to the location of rotated objects and obtains accurate features of rotated objects to avoid missing or false detections caused by object rotation.The experimental results show that this method can more accurately locate rotated objects.The mAP metrics are respectively improved by 0.89%and 0.76%compared to the DFFNet,and by 3.26%and 8.39%compared to the state-of-the-art methods on the DIOR-R and DOTA datasets.