Research On Single-stage Remote Sensing And Aerial Image Object Detection

Remote sensing and aerial image object detection is one of the important research directions in the field of computer vision,which is widely used in various fields such as urban management,resource allocation,natural disaster warning,ecological environment detection,etc.Since most remote sensing devices are shot from a bird’s-eye view,the targets in remote sensing and aerial images have an oblique angle.At the same time,the targets to be detected often have many objective factors that are unfavorable for detection,such as dense distribution,overlap,complex background,large size changes,large differences in quantity,wide aspect ratio range and so on.In order to overcome these difficulties and further improve the detection performance of existing object detection models on remote sensing and aerial images,this thesis studies based on a single-stage detection model in feature fusion and lightweight two directions.The main work content is as follows:1)In order to efficiently and accurately detect rotated objects in high-resolution remote sensing and aerial images,this thesis proposes a dual-channel feature fusion detection model.The model adopts a dual-channel downsampling structure,which uses convolution kernels of multiple sizes to filter the input image to obtain feature information from different perspectives.The coordinate attention module is restructured using the self-attention mechanism in Transformer to globally calculate the relevance between channel vectors,allowing the model to better denoise and enhance foreground information.A channel concatenation module is designed to extract information on channels from multiple dimensions,avoiding the loss of effective features.In addition,this thesis addresses the shortcomings of applying the Io U loss function to rotated object detection by introducing trigonometric functions to calculate angle loss,thereby improving the learning ability and convergence speed of the model.2)In response to the contradiction between the large number of parameters,high computational resource requirements of target detection models,and limited software and hardware resources of drones,satellites and other equipment,this thesis proposes a lightweight remote sensing and aerial image target detection model.The model features a multi-scale inverted residual module which significantly improves the feature extraction ability and model generalization ability.The cross-attention module is used to further fuse the feature maps obtained by the backbone network with highdimensional information obtained by down sampling,expanding the model’s receptive field and enhancing the feature representation of targets,allowing deep feature maps to simultaneously have accurate semantic information and texture information.Finally,an anchor-free detection head is used to predict suspicious objects,further reducing the computational complexity of the model and improving its detection speed.The experimental results demonstrate that the dual-channel feature fusion detection model achieved 77.7% of m AP on the DOTA-v1.0 dataset,showing significant improvements over other comparative models.The lightweight remote sensing and aerial image object detection model reduced the number of parameters and computational complexity by 66.7% and 59.0%,respectively,compared to the dual-channel feature fusion detection model,while only decreasing the detection accuracy by 5.4% on the DOTA-v1.0 dataset.The prediction accuracy on the HRSC2016 dataset reached 88.7,proving its high efficiency in detection capability.