Research On Multi-UAV Target Detection And Recognition Technology Based On Visible Light Image

In the field of computer vision research,target detection and recognition occupies a very critical position.In recent years,with the rapid development of artificial intelligence technology and the large-scale emergence of image data and video data,target detection has been widely used in the fields of target tracking,image analysis,and unmanned driving.This thesis uses two mainstream target detection algorithms based on saliency and deep learning to detect and recognize multi-UAV targets in visible light images.The hypercomplex frequency domain transform(HFT)model based on the salient target detection method uses the RGB characteristics to construct a quaternion function to detect the target in the frequency domain.The model has certain adaptability to multiUAV target detection under the sky background,and has excellent performance for small-scale target detection and target detail retention.In order to realize the detection of UAV targets in complex scenes,the Faster R-CNN target detection algorithm is studied.Through comparison experiments,Res Net50 with better detection effect is selected as the backbone network of the model,and a new model is constructed by fusing the Feature Pyramid Network(FPN)to effectively detect multi-scale targets,and uses feature enhancement modules and parameter modifications to improve model performance.Experiments have found that the improved model improves the detection capabilities of small-scale UAV targets.The research content of this thesis includes the following aspects:1.Research the saliency-based target detection method,and conduct an in-depth analysis of the hyper-complex frequency domain transform model,the model uses a multi-scale Gaussian smoothing template to make the edge information of the target lose.In order to solve this problem,use wavelet transform multi-scale decomposition of the amplitude spectrum image is performed,the best amplitude spectrum and phase spectrum are selected to generate the best saliency map,and the sky cloud boundary is removed to obtain the final saliency map.The improved model not only increases the detection speed of the model,but also retains the edge and detail features of humanmachine targets also enhance the model’s ability to detect small-scale UAV targets.2.Researched the Faster R-CNN target detection algorithm framework,constructed a UAV target detection data set with different flight attitudes in various scenarios and expanded the data set with image enhancement methods.In order to select the best backbone network to build the Faster-RCNN detection model,three backbone networks(VGG16,Res Net50,Res Net101)were selected as the Faster R-CNN model backbone network for model training,and the trained model was tested and found to be Res Net50 as The Faster R-CNN model of the backbone network has the highest detection accuracy,and Res Net50 is selected as the backbone network of the detection model.3.In order to improve the model’s ability to detect multi-scale UAV targets,the feature pyramid network was studied and merged with the Faster R-CNN model to construct a model framework for multi-UAV target detection.The network model was trained,and the test found that the new model has significantly improved the detection accuracy of UAVs.In order to further enhance the feature extraction capabilities of the network,three types of convolutions were integrated on the regional proposal network to extract features,and the anchor basic scale and anchor ratio were adjusted according to the characteristics of the drone,and the final drone detection model was constructed.Experiments found that the improved model further improves the detection accuracy of UAVs and can effectively detect multi-scale UAV targets.