Research On Aircraft Target Detection Technology In Remote Sensing Image Based On YOLO Algorithm

In the field of digital image processing,remote sensing image is an important spatial information resource,which is widely used in road defect detection,building extraction,resource exploration,crop disease and insect detection,urban planning,military reconnaissance and other fields because of its unique imaging perspective.In recent years,the efforts of researchers have made the remote sensing technology develop rapidly.However,at present,the acquisition ability and processing ability of remote sensing images are seriously out of balance,which makes a large number of remote sensing images cannot be effectively used in practical projects,resulting in a waste of resources.In the research field of remote sensing image processing,aircraft target is not only an important means of transportation,but also a military target on the battlefield,so its detection technology has an important research value.At present,the aircraft target detection algorithm based on traditional remote sensing image is difficult to adapt to the detection task under the complex background because of its poor flexibility and robustness.Therefore,how to quickly and accurately extract the aircraft target from the remote sensing image under the complex background has become one of the focuses of remote sensing image processing.In recent years,deep learning technology has shown great brilliance in the field of artificial intelligence and attracted extensive attention from researchers.Among them,many algorithms based on convolutional neural network have made many innovative progress in the field of image processing.Through experimental comparison and verification,based on the YOLO series algorithms commonly used in the engineering field,this paper carries out research on the application of YOLOv4-tiny algorithm in aircraft target detection of visible remote sensing images,focusing on the improvement of activation function in the algorithm and the improvement of the algorithm for complex background and small target detection.Specific research contents and innovation points are summarized as follows:(1)The activation function in the convolutional neural network mainly undertakes the task of nonlinear data processing,and its performance directly affects the accuracy of the algorithm.For remote sensing image plane target detection tasks,this article is based on YOLOv4-tiny algorithm,enumerate the activation function is commonly used in target detection algorithm,through a large number of experiments,and focus on different activation function in the training of the precision ratio,recall ratio and AP of the curve and test time,eventually selected training process stable and the most suitable for remote sensing image plane target detection task swish activation function,and apply the activation function can improve the recall rate of 3% and the AP of 0.19%.At the same time,since there is no open source remote sensing image data set dedicated to aircraft detection at present,the experiments in this paper all use the self-constructed remote sensing image aircraft target data set.(2)In the actual use of YOLOv4-tiny algorithm for aircraft target detection task,there is a complex background and small target aggregation problems resulting in difficult detection.For the first problem,using the scheme of improving the backbone network,this paper deeply studies the backbone network of YOLOv4-tiny algorithm,improves the original backbone network based on Darknet 53 network and CSPNet,and expands the input resolution of the network.Experimental results show that the improved algorithm has a good detection effect in the case of complex background interference and overexposure,the recall rate and AP are improved by 21% and 10.64%,respectively.Aiming at the problem of small target detection,this paper uses the spatial pyramid pooling layer mitigation algorithm to be sensitive to the scale of the target.Two types of spatial pyramid pooling layer are used to improve,and the optimal spatial pyramid layer is selected through experimental comparison,and the selected ones are tested.The sensitivity of the activation function to the two improvement schemes.In the final experimental verification and comparison,evaluation was made from subjective evaluation and objective evaluation.The accuracy,recall and precision of the improved YOLOv4-tiny algorithm increased by 1%,21% and 11.62% respectively.