Research On Small Object Detection And Fine-grained Recognition In Optical Remote Sensing Image

Object detection and fine-grained recognition in optical remote sensing images is a technology that determines spatial object position by analyzing satellite or aerial optical images and subdivides it into specific subcategories.Although there have been some attempts to solve the problem of remote sensing image object detection,there are still many difficulties:(1)There are various interference factors in optical remote sensing images which affect the detection performance.Optical remote sensing image has the characteristics of complex background and changeable environment,which brings a lot of useless features and causes serious interference to the extraction process of object features.Secondly,many densely arranged small objects in optical remote sensing images occupy very few pixels and interfere with each other,thus affecting the detection and recognition.(2)Fine-grained object recognition in remote sensing images is a challenging task.The orientation of objects in optical remote sensing images is different,and the misalignment of features makes the lack of detail features lost.In addition,due to the long observation distance of remote sensing images,the detailed texture features of the object are insufficient,and there is a lack of efficient extraction methods for detailed features.Therefore,because of the problems in remote sensing images such as too much interference information,dense distribution of small objects,and insufficient detail features,this paper proposes two solutions: remote sensing small object detection method based on the Fusion Cascade Attention Module(FCAM)and fine-grained recognition method based on alignment convolution and local feature learning.The two schemes are described as follows:(1)Considering the complex environment and the problem of the densely arranged small objects in optical remote sensing,a remote sensing small object detection method based on the FCAM is proposed.Firstly,to reduce the interference of complex background and weather environment on feature extraction,a multi-scale attention module was designed inspired by the attention mechanism.Secondly,considering the difficulty of small object feature extraction,the FCAM is designed to extract and fuse multi-scale features between different feature layers.Then,the loss function is optimized to balance the proportion of positive and negative samples in the representation method of rotating anchor frame with midpoint offset representation.Finally,comparative experiments are designed on several large-scale datasets in aerial images to verify the advantages of the proposed method in remote sensing image object detection,especially in small object detection.(2)In this paper,aiming at the problem of feature misalignment and insufficient fine-grained features in optical remote sensing images,we propose a fine-grained recognition and detection network for optical remote sensing objects based on aligned convolution and local feature learning.It can make end-to-end detection and classification.Firstly,the alignment convolution layer optimized for the rotating object is used to deal with the problem of misalignment between the convolution feature and the object.Then,a local feature progressive sampling network is designed to extract the local detail features.Finally,comparative experiments are designed on several large-scale datasets for object detection and fine-grained recognition in aerial images.Experimental results show that the improved method has a more vital ability of detail feature extraction and significantly improves the performance of fine-grained recognition.