Research On Image Recognition Based On Convolutional Neural Network In Optoelectronic System

An Optoelectronic system is a comprehensive system that integrates optical system,precision machinery,electronic system,computer system,and other systems.It has a broad and important application in the fields of tracking and aiming,astronomical observation,communication transmission,precision measurement,material processing,and so on.The function of signal processing is analyzing and processing different kinds of input information for subsequent control units,which is the core part in the optoelectronic system’s work.Image is one of the most widely used information in the system since it has the characteristics of rich information.Therefore,image recognition becomes the key problem in signal processing.In this thesis,the convolutional neural network(CNN)is applied in the image recognition tasks in the optoelectronic system,which is conducive to improving the operational efficiency of the optoelectronic system.With the development of unmanned aerial vehicle(UAV)technology,UAVs with rich functions are widely used in a variety of applications,which brings the favourable circumstances and challenges to the optoelectronic system.For our UAVs,the optoelectronic system can ensure their long-term and efficient operation by completing tasks such as laser wireless power and ground to air communication.For the UAVs of unknown source,the optoelectronic system shall have the ability to detect and monitor the target for security.The above applications include target mark and emission,in which image recognition plays an important role.For the former,the research object is UAV keypoint recognition,which can mark the energy conversion equipment or the signal receiving device on the target,and provide guidance for charging or communication beams.For the latter,the thesis aims at completing the task that predicting the wavefront phase from the far-field spot image,which can be used to combine high-quality beams for wireless power,signal transmission and so on.The main research contents are as follows:(1)A UAV keypoint regression algorithm based on the cascaded neural network:In order to suppress the adverse impact of interferences in the background on object keypoint recognition,we designed a two-stage UAV keypoint regression algorithm.In the first stage,we used the anchor-based object detection algorithm to obtain the position information of the target,crop the candidate regions from the original image and send them to the next stage.In the second stage,the algorithm locates the keypoint based on the global information of the target.The design of the cascade model is conducive to filtering background information.The lightweight structure improves the real-time performance of the model.(2)A UAV keypoint recognition algorithm based on multi-level feature integration and confidence prediction: In order to make sure that the model is able to locate multiple keypoints with the same classification on the target,this thesis proposes a keypoint recognition algorithm with heatmap representation.We designed a novel multi-level feature integration structure to improve the performance of the model on sensing context information and extracting the robust feature,so as to deal with the problem of self-occlusion and unclear local features of keypoints in the UAV target.Besides this,this thesis designs a confidence prediction mechanism to supervise the refine network and improve the performance of the model on the keypoints which are difficult to identify.(3)An intelligent piston phase prediction method: This thesis analyzed the mapping relationship between piston phase and far-field image in theory,determined the causes of phase ambiguity,and proposed the phase modulation method to solve this problem.Our solutions can prevent the models from diverging.Then the thesis established a CNN model based on phase modulation for phase recognition,and compared its performance with SPGD.The simulation results show that our model has advantages in phase compensation effect and overall running time.In addition,the thesis proposed a novel evaluation function and solved the problem of unreasonable evaluation.Finally,the thesis designed a novel loss function to effectively improve the training effect and generalization ability of the CNN model.The thesis completes representative tasks of image recognition for UAV wireless power transmission and UAV defence in optoelectronic system with using CNN.In the target-marking stage,our model can effectively locate the keypoints of UAV with high precision,which is helpful to track the specific parts of the target accurately.In the emission stage,our model can effectively solve the problem of phase ambiguity,and obtain accurate phase prediction through single-step inference,which is conducive to combining high-quality coherent beams efficiently.At the same time,the above researchs verify the effectiveness of image recognition based on CNN in solving the task of signal processing in the optoelectronic system,and lays a certain foundation for the application of CNN in the system.