Research On Detection Algorithm Of Complex Ground Objects In Optical Remote Sensing Images

Optical remote sensing image target detection has great research value for the development of remote sensing and aviation fields,and has a wide range of application scenarios in many practical scenarios such as ocean detection,ground urban planning and land monitoring.With the development of aviation technology and remote sensing technology,the number of optical remote sensing images that can be used for analysis has grown rapidly,so we need more efficient intelligent analysis tools to replace manual analysis.In recent years,with the enhancement of hardware computing power and the rapid development of deep learning,target detection methods based on convolutional neural networks have highlighted many advantages over traditional target detection methods.Therefore,according to the characteristics of optical remote sensing images,it is particularly important to propose a target detection method based on convolutional neural network suitable for optical remote sensing images.The work of this thesis is based on the research on the target detection method of optical remote sensing images based on deep learning,and conducts in-depth research on the multi-scale problem of targets in optical remote sensing images,the complex background information of targets and the diverse characteristics of target shapes.The contributions of this thesis are summarized as follows:（1）Aiming at the characteristics of multi-scale targets and complex target backgrounds in optical remote sensing images,a target detection method based on attention mechanism is proposed.First,the global context module is integrated into the feature extraction network,so that the feature extraction network can selectively increase the weight of the target feature during feature extraction,suppress the interference of background noise,and strengthen the feature extraction ability.Then,the coordinate attention mechanism impro ves the feature fusion method of the feature pyramid,so that the target detection network can achieve better performance on the target multi-scale problem.Finally,according to the difference between the label classification task and the bounding box reg ression task,the corresponding information of the label classification branch and the bounding box regression branch is given to further improve the accuracy of the target detection network.（2）Aiming at the characteristics of diverse target shapes and w ide distribution of aspect ratios in optical remote sensing images,a remote sensing image target detection method based on pixel point regression is proposed.Using a regional recommendation network without anchor boxes to replace the regional recommendation network based on preset anchor boxes avoids the setting of the hyperparameters of the preset anchor boxes,and the preset anchor boxes cannot cope with the problem of target shape diversity well.Giving image features to guide the construction of anchor boxes can better cope with the characteristics of target shape diversity and aspect ratio distribution.In addition,the training strategy of dynamic threshold is used.With the continuous improvement of the performance of the target detector in the training process,the threshold value of positive and negative sample allocation is gradually increased,so as to obtain a higher quality target detector.