Research On Aircraft Target Classification Algorithm Based On Feature Fusion Convolutional Neural Network

Whether in military or civilian use,UAVs have been widely used.When flying at a low altitude,UAVs can quickly and accurately classify and identify obstacles ahead,which is an important guarantee for UAVs to complete automatical avoidance and reduce air traffic accidents.Compared with other target classifications,the low-altitude aircraft target classification not only needs to complete conduct the cross-species classification,but also completes the classification of different subtypes of aircraft sub-categories,and there are many types of obstacles.At the same time,due to light,air background,and other external factors.The influence of these factors has brought difficulty to the classification problem.Therefore,the specific methods to solve the UAV target classification problem in this paper are as follows:Firstly,because the data set plays a vital role in the training and optimizing of the network,in response to the current lack of open source low-altitude aircraft target classification data sets,this article has realized the collection and production of the low-altitude aircraft target classification LAAT data set.For the problem of data quality and imbalance,image preprocessing,image rotation,flip and random clipping are used to enhance and expand the data.Secondly,the convolutional neural network contains many specific subnetwork structural frameworks,and each framework has its unique advantages.As the LAAT data set includes cross-species classification and aircraft sub-class classification,a three-layer modular convolutional neural network MCNN was proposed,and the LAAT data set and FGVC-Aircraft data set were used for simulation verification.Thirdly,due to the small differences between aircraft sub-categories,the classification accuracy of convolutional neural networks is low,and the high-order features extracted by CNN lack global rotation invariance.Therefore,this paper proposes a convolutional neural network model based on feature fusion.First of all,the Canny operator and the improved G-LBP operator are selected to extract low-level contour and texture features.The second,the feature fusion method of the add was used to fuse the two extracted low-level features into lower-order features.Then,the concat feature fusion method was used to merge the extracted low-order features and high-order features into complementary composite features with global rotation invariance for classification,and the LAAT data set and the FGVC-Aircraft data set were used for experimental verification.Finally,it is verified by experimental simulation that the method proposed in the article can solve the conventional problems encountered in UAV target classification.