Aerial Image Target Detection Algorithm Based On Deep Learning

With the increasing use of unmanned aerial vehicles and the development of aerial photography technology,the algorithm requirements for automatic data extraction of aerial images of unmanned aerial vehicles continue to increase,and the research of target detection algorithms for aerial images of unmanned aerial vehicles is of great significance.The image features of aerial images are very different from those of general images.The details and features that are easy to be lost have an error size in recognition.The height of the air flight is different,and the angle of view and the size of the target are also different.The same target will have different characteristics and details,which will increase the detection defects.Targeted research and experiments on aerial images are needed.First,for drone aerial images,the main research is a high-speed detection model for small objects.By improving the Cascade R-CNN of multi-scale fusion,a detection method for drone aerial images is proposed.Compared with the traditional method,paper uses the idea of merging feature layers,and merges feature maps at different levels to improve the algorithm’s ability to detect small targets.Through testing,this method has achieved good results on the Vis Drone 2019 data set.Afterwards,for the improved Cascade R-CNN network,the model quantization and pruning strategies were studied and improved.Based on the experimental results,the advantages and disadvantages of each method were compared,and a model lightweight method suitable for aerial image target detection problems was generated.Compare the detection accuracy of the model in float 32,float16,int 8 data format,design an improved BN pruning method,cut the feature layer with the least impact on the network by the resolution coefficient of the BN layer,and use the mixed weight The strategy makes the method have good effect in the complex detection model.Finally,in order to achieve rapid deployment of the model,around the drone image,under the open source image annotation software Label Me,through the Polygon RNN and Mask R-CNN networks pre-trained on the standard datasets,automatic annotation of the drone image and artificial Aid to modify and adjust.Use the collected data and public data sets to test the automatic labeling software,and fine-tune the existing network through the labeled data to detect the performance improvement effect.