Object Detection In Remote Sensing Images Based On Background Information Integration

Remote sensing information processing plays an important role in many fields such as environmental protection,urban planning,and military reconnaissance.With the continuous advancement of satellite technology,the number of high-quality remote sensing images has increased significantly,which provides great convenience for the direction of remote sensing image processing.However,due to the characteristics of high-resolution remote sensing images with various scales,rotating objects,dense distribution,occlusion,and complex backgrounds,algorithms migrated from natural scene target detection cannot achieve excellent detection performance.In addition,the current mainstream remote sensing image target detection algorithms only focus on the characteristics of the object and ignore background information such as scene information and common sense knowledge.In order to improve the accuracy of remote sensing target detection and integrate background information into the algorithm,this paper proposes related solutions.(1)For the problem of fusing scene information,the method of Ro I Transformer is improved.The Scene Module is introduced to combine the scene information with the extracted object features.In this chapter,on the basis of Faster RCNN and Ro I Transformer,the scene information propagation mechanism of parallel to Ro I Transformer is designed.At present,the mainstream remote sensing image target detection algorithm has achieved excellent performance in extracting object features,but the mining of scene features such as context features is insufficient.Especially in the face of changing scales,dense distribution and occlusion,scene information such as context features has a considerable positive impact on the reasoning process of the algorithm.Through a large number of experiments,the improved algorithm of Ro I Transformer proposed in this paper has achieved better results than the baseline model in terms of detection performance.(2)In terms of expression of background knowledge,the co-occurrence probability between each category and the co-occurrence probability between each category and waters is obtained through the statistics and analysis of public remote sensing datasets.In order to transform the obtained knowledge in the form of probability into the knowledge fusion module suitable for this paper,this paper improves the existing conversion method,and performs special treatment for 0probability(non-co-occurrence category),and changes the output value of the original formula domain,the method can represent the probability size more effectively.(3)For the problem of fusing background knowledge,an improved method based on Oriented RCNN is proposed.In order to make use of the background knowledge of digitization,this paper proposes the Knowledge Inferecnce Module,and adopts the residual structure to ensure that the detection performance does not degrade.Through a large number of experiments,the improved algorithm based on Oriented RCNN proposed in this paper has achieved higher mean average precision compared with the baseline model on DOTA and DIOR datasets while ensuring that the inference speed is basically unchanged.