Detection Of Small Objects In Remote Sensing Images Based On FPN And Cascade Structure

With the development of deep learning,object detection has made significant progress in many real-world scenarios.Despite this remarkable progress,there are still huge challenges in small object detection,especially in the field of remote sensing images.Due to the particularity of imaging,remote sensing images usually contain a large number of small objects.However,small object detection has a pivotal significance no matter in the civil or military fields.Research has found that there are four main problems in remote sensing image small object detection: weak representation ability caused by few effective features,the feature of small object is easy to lose,small object is easy to be submerged in a complex background,and small object is difficult for positioning.This thesis mainly proposes corresponding improvement methods for these problems.The main work consists of four parts:(1)In view of the weak representation ability of small targets in remote sensing images and the problem that small object is easy to lose its features in the deep network,we propose a novel feature pyramid network——Refine FPN,and verify it on multiple remote sensing image data sets.Compared with the original FPN,Refine FPN uses the resize-conv upsampling operation,redesigned the building block for deep and shallow feature fusion,and added skipping connections between the input and output of the same level.(2)Aiming at the problem of few effective features and weak representation ability of small objects in remote sensing images,we propose a cross-layer attention network,and verify it on multiple remote sensing image data sets.Inspired by the human attention mechanism that recognizes long-distance small objects,we designed the cross-layer attention network.In the cross-layer attention network,we simulated the attention mechanism by designing an improved non-local attention module,and then added a cross-layer integration and balance module to provide stronger and more balanced features for subsequent detection network.(3)In order to solve the problem of difficult localization of small objects in remote sensing images,we proposed a network for small object detection remote sensing images based on context and cascade structure,and verified it on multiple remote sensing image data sets.The proposed object detection network mainly includes two parts: context transfer module and cascade detection network.Among them,the context transfer module is to transfer context information to each Ro I generated in the RPN stage to assist in object localization,while the cascaded detection network contains a sequence of detection networks with a set of increasing Io U thresholds.The cascaded detection network can solve the over-fitting problem and suppress the false positive samples in detection.(4)Considering that the lack of data sets for small object detection in remote sensing images,we analyzed the DOTA 1.5 version data set,and composed a data set containing a large number of small objects and covering four categories,named Small-DOTA data set.This data set can be used to evaluate the model’s ability to deal with the scenarios where there are many small objects.In addition,we preprocessed the Small-DOTA dataset and the OHDSJTU-S dataset by sliding window,and transformed the ground truth of the OHD-SJTU-S dataset.