Research On Rotated Object Detection Algorithm For Aerial Images Based On Feature Focusing

Nowadays,with the rapid development of high-altitude imaging technologies such as drones and satellites,high-resolution aerial images are widely used in military strategy,resource exploration,urban planning,traffic control and many other fields,profoundly affecting social production and human lives.In practical applications,it is usually necessary to use specific means to extract,process and analyze useful information in aerial images.Among them,object detection of aerial images aims to extract specific objects from complex aerial images,which is the most commonly used data processing method for aerial images.With the development of deep learning and big data technology,many object detectors based on deep learning can complete high-performance in common scenarios.However,aerial images often have the characteristics of high resolution,complex texture and noise information,large differences in object scale and aspect ratio,and rotated objects,which brings great challenges to conventional object detectors.Researchers need to modify the functional components of the detector,and design a specific feature encoding module for the network model to complete effective detection of aerial image objects.This paper is dedicated to the in-depth study of aerial image object detection,and proposes a feature-focused aerial image rotating object detection algorithm,which is named Align-Focus Detector(AFDet).AFDet is composed of feature extraction backbone network optimized by window self-attention mechanism,path aggregation multi-scale feature fusion neck network,deep feature alignment focusing module and rotating object detection decoder.It can achieve excellent detection tasks for complex rotating objects in aerial images.performance.The main work of this paper is as follows:(1)This paper optimizes the feature extraction backbone network and multi-scale feature fusion network of the detector for the challenges of high-resolution aerial images,complex texture information,and large differences in object scale and aspect ratio.First,the window self-attention mechanism is used to optimize the deep feature extraction backbone network,which fully combines the local feature encoding ability of convolutional structures and the large-scale global feature interaction ability of self-attention mechanism,which can finish more effective feature extraction for aerial images.Secondly,using path aggregation network as the multi-scale feature fusion neck network,the forward fusion path and the reverse fusion path can provide effective information transmission between deep and shallow features,which can effectively improve the recognition ability of multi-scale objects in aerial images.(2)In view of the inconsistency between the horizontal receptive field with the spatial range of the rotated object feature and the lack of feature adaptive drive in the optimization coding of convolutional feature,this paper designs an Align-Focus Module(AFM)for feature of each level.AFM first predicts an adaptive-rotated anchor box through a lightweight fully convolutional structure for each feature point.According to the adaptive-rotated anchor box,the receptive field of the point feature is aligned with the anchor box by using aligned convolution to generate adaptivelyaligned feature.Finally,the adaptively-aligned features and input features are adaptively interacted with spatial information through an aligned multi-head attention mechanism,and the coding response of point features are focused on the effective texture regions within the aligned receptive field and generate aligned-focused features.In addition,for the rotated object detection decoder,active rotation filter convolution is used to generate direction-sensitive features for object location,and for classification prediction,direction-response maximum pooling is used to further generate direction-invariant features to reduce feature redundancy and enhanced features that contain object category information.Finally,for the boundary problem caused by the arbitrary rotated angle of the aerial image object,the KF-IOU loss function is used to supervise the object location output,so that the model can be trained more effectively,which can further improve the precision performance.(3)This paper uses the public DOTA dataset and HRSC2016 dataset to carry out necessary verification experiments.By analyzing the training process and accuracy verification results recorded during the experiment,it is proved that the innovations and modifications proposed in this paper have great impact on effectiveness in improving object detection performance in aerial images.In addition,in the comparison experiments,the accuracy performance of AFDet is further compared with other excellent aerial image object detectors.Finally,this paper also compares AFDet with the benchmark detectors in visual experiments.AFDet solves various deficiencies of the benchmark model,including misalignment between the detection bounding box and the object,missed detection,false detection,and incomplete detection.Through a series of experiments,it is fully proved that AFDet has excellent performance in the task of aerial image object detection.