Research On Efficient Detection Of Multi-scale Objects In Optical Remote Sensing Images

With the rapid development of remote sensing technology,remote sensing image object detection has become an important research direction in the field of computer vision.In recent years,remote sensing image object detection algorithms based on deep learning have been greatly developed and widely used in fields such as land monitoring,resource investigation and precision agriculture.However,the current optical remote sensing images have the problems of poor object detection and large size of network models.In response to these problems,this thesis carries out the research of efficient detection methods for multi-scale objects in optical remote sensing images,and the innovative research results and the main completed works are as follows:(1)A joint multiscale and attention mechanism object detection algorithm for remote sensing images is proposed.Aiming at the problem that the complex background of remote sensing images and the large difference of object scales lead to the poor detection effect of existing algorithms,the remote sensing image object detection algorithm with combined multiscale and attention mechanism is proposed.Firstly,the void space pyramid pooling module is improved to increase the perceptual field of different size images.Secondly,the attention module is proposed to learn feature map channel information and spatial location information to improve the feature extraction ability of the algorithm for remote sensing image object regions in complex backgrounds.Then,the weighted bidirectional feature pyramid network structure is combined with the backbone network to enhance the fusion of multi-level features.Finally,a distance-based non-maximum suppression method is used for post-processing to improve the problem of easy overlapping of detection frames.The experimental results on DIOR and NWPU VHR-10 datasets show that the proposed algorithm achieves an average accuracy of 71.6 % and 91.6 %,respectively,which is 2.9 % and1.5 % better than the mainstream YOLOv5s algorithm,respectively,and is more advantageous in complex remote sensing image object detection tasks.(2)An all-round remote sensing image object detection algorithm based on feature fusion and angle classification is proposed.Aiming at the problem that the existing detection algorithms are difficult to accurately locate the objects because of the dense arrangement and different directions of the objects in remote sensing images,this thesis proposes an all-round remote sensing image object detection algorithm based on feature fusion and angle classification.First,in order to improve the feature extraction capability,a pure convolutional model is introduced as the backbone network,and the Neck layer of the network model is redesigned to reduce the number of model parameters.Secondly,an enhanced connected feature pyramid network is proposed,and the lateral connection part for deep and shallow layer feature fusion is redesigned,and a jump connection is added between the input and output of the same level feature map to enrich the feature semantic information.Then,the angle prediction branch is introduced,and the angle regression problem is transformed into a classification problem using the loop smoothing labeling method,which solves the problem of abrupt changes in the boundaries of the rotation frame while realizing the object frame rotation.Finally,a variable parameter is added to the original localization loss function to meet the accuracy of the bounding box regression under different IoU thresholds,so as to obtain more accurate object detection results.The experimental results on the directional remote sensing object datasets HRSC2016 and DOTA show that this algorithm achieves better detection results compared with other advanced algorithms.(3)The lightweight remote sensing image object detection algorithm without anchor frame is proposed.A lightweight remote sensing image object detection algorithm without anchor frame is proposed to address the problems of large number of parameters in the existing remote sensing image object detection model in practical applications and the difficulty of deployment in mobile devices.Firstly,the DWS-Sandglass lightweight module is designed to reduce the model size.Secondly,the model activation function is improved to ensure the detection accuracy.Then the redundant channels of the anchorless frame algorithm are pruned to reduce the number of model parameters,and finally the parameter-free attention module SimAM is introduced to enable the network to focus on more important feature information.Experimental results on the HRSC2016 dataset show that the algorithm is faster in detection with comparable detection accuracy and smaller in model size than the current mainstream anchorless frame detection algorithm,which is more suitable for deployment in mobile devices.