Research On Detection Of Camouflage Target In Satellite Visible Light Image And Algorithm Optimization

With the continuous development of deep learning technology,the accuracy and detection speed of specific target detection have been greatly improved,and the target detection technology for military detection has gradually landed.Camouflage is often used to protect one's own combat units.It can make the reflected light wave of the camouflage target similar to the light wave emitted by the surrounding scenery,thus achieving the effect of confusing the enemy.Whether the camouflage corresponding to the target detection technology used on satellite images can be confusing is a safety issue worth studying.This paper collects relevant satellite image data and uses the latest target detection techniques to design experiments related to camouflage camouflage research.In the experiment we proposed a new network model to increase the validity of the experimental conclusions.Through a series of experiments and analysis of the data,we proved that the camouflage camouflage effect does not play a role in the target detection.At the same time,this paper optimizes the target detection related algorithm.The optimization includes four aspects:high availability,speed improvement,high cost performance and accuracy improvement.In the research of high availability,this paper designs a new learning goal for the target detection network,so that the model can meet the expected detection results for different sizes of input.The network modified according to this method can accept multi-size input images.It is beneficial to control the computing power required,and it is convenient for the model to be transplanted on different platforms,thus improving the model availability.In the research of speed improvement,this paper studies the elimination of the repeating frame stage of the target detection.After analyzing the calculation process of the intersection ratio,through scientific modeling,it analyzes the frame for a known size.The distribution of the boxes of unknown size up to the specified cross ratio must be limited to a range and the size of the known box is established as a function of the range.Based on this function relationship,this paper proposes a new non-maximum value suppression algorithm,which reduces the number of calculations of the intersection ratio based on the intersection and distribution information of the box.Compared with the traditional non-maximum suppression algorithm,the algorithm The time complexity can be reduced from O(9)~2)to O(9)),greatly reducing the time consuming to eliminate the repeating box phase.In the cost-effective research,this paper investigates a variety of lightweight network modules and accuracy improvement modules,and validates their effects on the model through design module replacement experiments.At the same time,through the combination of modules,a set of accuracy is maintained.The solution,which is time-consuming and halved,effectively improves the cost performance of the network.In the research of accuracy improvement,this paper analyzes the problems existing in the processing of output boxes by traditional methods,and proposes an adaptive confidence threshold method to process the final output.By using the machine learning related method to model the performance distribution of the model,the accuracy of the prediction is improved,and the workload of the human selection threshold is effectively reduced.