Context Aware CNN For Object Detection From VHR Remote Sensing Imagery

With the development of aerospace technology,the resolution of remote sensing images has been continuously improved,and target recognition of remote sensing images has a wide range of applications in many fields such as defense security,traffic monitoring,urban management,intelligent video surveillance,content-based image retrieval,robot navigation,and augmented reality.As one of the basic research topics in the visual field,target detection is the basis and prerequisite for many advanced visual tasks,but overall it is still subject to the following three factors:the image quality is not high,the target itself is too small,and the detection method is imperfect.The poor image quality is due to the fact that remote sensing images are easily affected by the natural environment such as sunlight,clouds,fog,and waves during the imaging process.At the same time,some targets further increase the difficulty of target detection due to their own small scale and the interference of peripheral noise.How to accurately extract the target and information of interest from a large number of complex scenes is of great significance.Rich feature representation is the key to target detection.In this paper,a context aware CNN for object detection from VHR imagery is proposed,which can divided into three aspect:scale context extraction,spatial context feature extraction,and feature fusion.This paper starts with the two kinds of contextual information in scale and space,and deeply studies the feature expression of the ground features.Besides,a more effective feature fusion method to integrate scale features and spatial features is proposed.The experimental results show that the proposed method can significantly improve the detection accuracy of the target and has more accurate results than the conventional target detection method.The article mainly includes the following:(1)The target detection method of remote sensing image based on scale context is studied.In this paper,the scale context is extracted based on deep learning.To solve the loss of detail information caused by convolution,we consider extracting the feature of target on feature maps of multiple levels,and fuse them for the target detection.First,feature extraction is performed on feature maps of different levels,and the effects of different hierarchical features on target detection are compared through experiments.Then,the extracted features of different scales are further merged to make up for the semantic gap between low-level features and high-level features,and to improve the accuracy of target detection.(2)The spatial context-based target detection method of remote sensing image is studied.The object always appears in a specific scene.The internal connection between the target and the surrounding environment can effectively reduce the uncertainty of the target detection and improve the accuracy of detection.Therefore,this paper focuses on the contribution and influence of spatial context on target detection.Through the experiment and analysis of the influence of different contexts and sizes on the target detection,an effective spatial context fusion method is proposed.The results show that the spatial contextf fusion helps improve the detection accuracy.(3)The feature fusion method is studied.Based on the previous research,the scale context and the spatial context is combined for target detection.A new and more effective feature fusion method is proposed for context feature fusion.The result proves that the improved feature fusion method can further improve the accuracy of the target detection.This article aims to improve the accuracy of high-resolution target detection.Through the study of feature expression and fusion,a context aware CNN for target detection from VHR imagery is proposed.The experimental results show that this method has a significant improvement on most target detection tasks.