| Due to the continuous advancement of urbanization and the rapid development of intelligent urban management,in-depth analysis of remote sensing images covering large areas and multiple targets is crucial.With the advancement of artificial intelligence technology and the update of computing devices,deep learning-based detection methods are gradually applied to remote sensing image analysis tasks,and they often have higher detection accuracy and anti-interference capability compared with traditional detection methods.As the main object of urban traffic management,vehicles have the characteristics of small size,large number,unbalanced number of categories,and complex surrounding environment in remote sensing images.It is of great significance to detect their location and number.The traditional detection algorithm based on target frame takes one candidate area as the detection unit,and the accuracy of the algorithm will decrease significantly when detecting smallsized and partially occluded targets.In addition,due to the limitation of the amount of data in a single dataset,it is difficult for the model to extract enough features through training,which limits the detection accuracy that the algorithm can ultimately achieve.As the main object of land resource management,the area occupied by buildings often changes,and it is a very important task to detect the changes.Change detection requires processing images taken at the same location and at different times.Because of the different lighting and seasons at the time of shooting,these images will show different styles,making the detected areas missing or noisy.For the characteristics of large image scale and complex environmental information,traditional methods often use visual attention mechanisms to optimize the features extracted by the model,but they optimize mostly different features or regions within a single image,without targeted optimization based on the differences between images.In view of the above problems,this paper improves the traditional methods in terms of detection form,model structure,data processing methods,etc.The main research work and innovations are as follows:1.The vehicle detection algorithm based on pre-training and random initialization fusion network for remote sensing images is designed: using the method of regression vehicle spatial density map to detect vehicles in pixel units,and using the clustering algorithm on the density map to locate and count the vehicles;designing the pre-training and random initialization fusion network,using the random initialization and pre-training models in parallel form together as encoders;the random initialization network is retained to fit the features of the current database,while the pre-training models are added to enrich the variety of features;the fusion feature encoder is designed to fuse the image features extracted from different classes of models,and the fusion of features from different classes of encoders is achieved by stitching and weighting the features.2.The remote sensing image building change detection algorithm based on the image difference attention network is designed: the feature difference map is obtained by using the pre-training model and the difference algorithm,and then the image features extracted in the network are optimized through the designed feature difference attention module,so that the network pays more attention to the area with obvious difference between the two images;uses the edge detection operator and the difference algorithm to obtain the edge difference map,and then uses the designed edge compensation module to compensate the edge of the image features extracted in the network,so that the edge of the region in the change detection result is more complete;an image difference attention network with an encoder-decoder structure is designed,two images are input,image features are extracted through the backbone,and the feature difference attention module and edge compensation module are used to optimize the features extracted by the network,and finally the change areas of the two images are obtained.Based on multiple public datasets,this paper firstly verifies the effectiveness of the above innovative methods through ablation experiments,and then through comparative experiments prove that the method proposed in this paper is superior to the traditional methods in relevant mainstream evaluation indicators.Finally,this paper summarizes the contributions of the proposed innovations to related fields and gives an outlook on the development of follow-up research. |
