| With the gradual increase of the proportion of UAV in military and civil use,the proportion of accidents has been increased to a certain extent.There are certain potential safety hazards in the UAV flight area.Therefore,how to improve the recognition accuracy of UAV has become one of the hot research directions.Previously,improving the recognition effect of UAV is optimized from the algorithm side after image input.At present,there is no research conclusion in academic papers that improving the image quality can improve the recognition accuracy of UAV.In fact,there is room to improve the recognition accuracy of UAV in each step from image acquisition to final recognition.A large number of academic reports have proved that the fusion of infrared image and visible image can improve the image quality and provide better pixel information for subsequent image-based experiments.However,it has not been proved whether it can provide more detailed information for UAVs with small targets,In order to further improve the efficiency of UAV image acquisition and infrared image recognition,the research of this thesis is based on the image fusion process of UAV.The main contents of this paper are as follows:（1）Image acquisition : Set up binocular camera and shoot data set and test set.By studying the common natural environment of UAV and the imaging characteristics of UAV,the imaging quality and image information of infrared and visible images of the UAV are studied,and a binocular camera is built for photographing.Simulate the real UAV scene,take 2565 UAV images with different complex backgrounds as the training set,498 visible image test sets and 498 infrared image test sets.（2）Image preprocessing: Preprocess the image.When studying and identifying small UAV,it is found that the target volume of UAV is small,so study need more edge information to highlight UAV,Laplace is used to preprocess 498 infrared image tests,enhance the edge information of the image,further improve the detail information of the image,and prepare for image registration.（3）Image registration: Image registration of visible and infrared images.Study the common methods and application scope of image registration,study the image characteristics of infrared and visible images,and creatively use the fusion registration method of Canny edge detection and ORB feature detection to complete the registration of498 visible and 498 infrared images,and prepare for image fusion at the same time.（4）Image fusion: The infrared image and visible image are fused.This paper studies the common methods and application scope of image fusion,analyzes the image characteristics of infrared image and visible image,carries out pixel level fusion through Harr wavelet,Keep the pixel information of the target object as much as possible,completes the fusion of498 visible image and 498 infrared image,and obtains 498 fused images.The comparison of relevant evaluation indexes before and after fusion proves that the quality of fused image is better,and is ready for target recognition.（5）Target recognition: Recognize the UAV image.The characteristics of one stage and two stage target recognition models are studied.Due to the small target characteristics of UAV,yolov3,which performs better in small target recognition in one stage,is selected for training framework construction,pre training and recognition.The experimental results show that 498 visible images before fusion and 498 fused images after fusion are tested respectively.The former map is 92.45% and the latter map is93.39%,which is increased by 0.94%.It shows that the method of fusing visible images and infrared images before UAV recognition is correct to increase image quality and Recognition accuracy. |
