Research On Remote Sensing Target Detection Algorithm Based On Attention Mechanism And Deep Network

With the rapid development of remote sensing observation technology,remote sensing target detection has become a hot topic at present,and the application field of remote sensing images has become more and more extensive.In recent years,high-resolution optical remote sensing images have attracted extensive attention due to their advantages of intuitively expressing target feature information and strong anti-interference ability.Due to the large difference in target size,serious background interference and dense distribution in remote sensing images,traditional target detection methods are not ideal for target detection in remote sensing images,resulting in lower detection accuracy.Therefore,in order to make up for the shortcomings of traditional methods,target detection methods based on deep learning have been proposed and widely used.In order to improve the shortcomings of existing remote sensing target detection and improve detection accuracy,this paper proposes a remote sensing target based on attention mechanism and deep network.detection algorithm.The main contributions of this paper are as follows:(1)Build a fusion attention mechanism.It mainly integrates channel attention and spatial attention,and combines them in an orderly manner.The channel attention mechanism is in the front,and the spatial attention mechanism is in the back to form a fusion attention mechanism.The image is first processed by the channel attention mechanism and then by the spatial attention mechanism,and then the final feature map is output.Adding a fusion attention mechanism to each residual block in the backbone network can enable the network to focus on the target area of the image,weaken the background interference,narrow the search range,and improve the detection and recognition ability of the target.(2)Build a multipath aggregate sampling network.The improvement of CSPDarknet and path aggregation network enables it to learn complex data transformation forms and improve the nonlinear expression ability of the model;in order to improve the adaptability of the detector,the information of the four feature layers of the backbone network is extracted,and the network up and down sampling of the image is increased.The process can enhance the semantic information of the image,and generate the thumbnail of the image to reduce the positioning error.(3)Propose smooth labeling and data augmentation techniques.Adding smooth labeling technology during data training can effectively prevent the training model from relying too much on label information,increase the distance between different objects,avoid overfitting in the training model,and effectively improve the generalization ability of the model.Using data enhancement technology can change the color,position,size and splicing method of the input image,making the training data more diverse and improving the stability of the algorithm.Thesis,the improvement of the above method is carried out and experimental verification is carried out on two public data sets.The overall average accuracy of the algorithm in the DIOR data set and the RSOD data set is 0.706 and 0.937,respectively,which are 0.054 and 0.04 higher than the original algorithm.The results demonstrate the effectiveness of this method.