| The object detection of optical remote sensing images has important research significance and value in the civil and military fields.In recent years,with the rapid development of artificial intelligence technology,the object detection method of optical remote sensing images based on deep learning has been widely studied and applied.Compared with natural images,optical remote sensing images have the characteristics of complex background information,huge changes in object scale,many small objects with dense arrangements,and variable object directions.These factors all pose a challenge to the improvement of the accuracy of object detectors for optical remote sensing images.This thesis conducts in-depth research on this background.The main research contents are as follows:Aiming at the situation that there are many small-scale objects and dense arrangements in optical remote sensing images,this thesis makes improvements based on SSD and proposes an object detection method based on feature fusion and residual attention mechanism.The model gives more semantic information to small-scale objects by fusing small-scale objects and their contextual features.By applying the residual attention mechanism to the extracted features,unnecessary information irrelevant to small-scale objects is suppressed,the interference of background noise on features is reduced,and the robustness and discrimination of features extracted by the model are improved.For the common problem of sample imbalance in object detection tasks,the focus loss function is used to replace the original sample allocation strategy to help the model dynamically control the weight of samples with different classification performances during training.Second,by combining factors such as training efficiency,center point distance,and aspect ratio,a new metric is used to measure the degree of overlap between the prediction bounding box and the ground truth box.Finally,to deal with the dense arrangement of objects,the non-maximum suppression algorithm is improved based on the distance between the center points of different bounding boxes.Through experiments,it is proved that the overall model and the improvement measures are effective in improving the accuracy,and an m AP of 71.1% is achieved on the DIOR dataset.Aiming at the problem that the directions of objects contained in optical remote sensing images have great uncertainty,this thesis proposes a two-stage anchor-free oriented object detection method based on polar coordinates.Given the instability of the model and the difficulty of training when the traditional Cartesian coordinate system represents the oriented bounding box,a five-parameter polar coordinate representation is used to model the objects.In terms of model feature extraction,due to the incompatibility between the receptive field and the input image resolution in the existing FPN,and the lack of communication of receptive field information at different scales,dense connections with different dilated rates are introduced based on the original network,which can obtain contextual information from different receptive fields and merges them.For the redundant information in the extracted features,the selfattention mechanism is used for filtering,which improves the feature discrimination ability of the model and strengthens the feature representation.Finally,simply and efficiently,highquality oriented region proposals are generated without presetting anchor boxes.The experimental results show that the overall model can achieve higher detection accuracy than other oriented object detection methods,and the value of m AP on the DIOR-R dataset reaches67.2%. |
