Multi-level Visual Perception Based Object Detection In Remote Sensing Imagery

In recent years,with the increasing development of the remote sensing technology,the quality of high-resolution remote sensing imagery has been gradually improved.As the primary data source of several practical applications,such as urban road construction,natural disaster management,underground resource exploration and so on,the object detection task in remote sensing imagery is particularly important.To be specific,the object detection task in remote sensing imagery is to accurately locate and classify the category-specific objects in a given aerial or remote sensing image,via some relevant algorithms in the field of pattern recognition or machine learning.However,due to the otherness in aspects of altitude and posture when remote sensing satellites are recording the images,objects in aerial images usually own some specific characteristics of their own when compared with those in natural images,just like the arbitrariness of arrangement orientation,disparity in scales and the high complexity of background.Because of these characteristics,object detection task in remote sensing imagery is technically challenging.To deal with the tough task mentioned above,this work improves the baseline method from three different perspectives,and achieves satisfying results on both horizontal and oriented bounding boxes tasks of three datasets including DOTA,UCAS＿AOD and NWPU VHR-10.The main works of this thesis comprise the following three ways:(1)We propose a feature attention based object detection algorithm in remote sensing imagery.To dispose some inevitable problems in remote sensing images,such as the excessive complexity of background,huge variation in the aspect ratios and scales of objects,we utilize different attention mechanisms in multiple stages of detection process which includes feature extraction stage,region proposal network and region of interest network.With the help of these multi-level attention mechanisms,the detection model can focus on learning these intrinsic representations from multiple aspects in an end-to-end framework,and then boost the detection results in remote sensing imagery greatly.(2)We propose a novel object detection method in remote sensing imagery with the guidance of soft semantic segmentation.Considering the character that skew bounding boxes in remote sensing images own a better cover of objects,we generate the pseudo box-wise masks on the basis of these oriented bounding boxes annotations.Then we take these masks as the auxiliary supervisions of detection model so that it can pay more attention to the foreground part of features.In order to enhance the objects’ feature representation and further improve the detection results of remote sensing imagery,we next make use of the auxiliary semantic feature produced by the segmentation module,multi-level convolutional features as well as the apparent characteristics for feature aggregation.(3)We propose a cascade object detection algorithm in remote sensing imagery.With the idea of multi-model cascading,we interleave multiple region of interest networks in series according to different annotation modes,so that we can utilize the peculiarities of bounding boxes with different forms.In the second stage,we use the skew bounding boxes as the main input of region of interest network,and make secondary predictions for both horizontal and oriented bounding boxes based on the skew boxes generated in the previous stage.By combing the advantages of horizontal and oriented boxes,the final cascaded model has the better ability to predict both horizontal and oriented bounding boxes,thereby improving the final detection results.In this thesis,based on the related characteristics of objects in remote sensing images,we make some targeted improvements to the baseline model from diverse aspects,our proposed algorithms obtain the state-of-the-art results on three public remote sensing object detection datasets.