Research On Aerial Target Recognition Algorithm Based On Deep Convolutional Neural Network

Target recognition is a branch of computer vision and pattern recognition,and it's an important part of artificial intelligence.Target recognition methods are divided into digital image processing method and based on convolution neural network method.Traditional digital image processing requires people to pickup many kinds of features,such as SIFT,HOG and LBP,which cannot express more abstract target category information.The method based on convolutional neural network can accomplish the abstract tasks of target classification,segmentation and scene interpretation,just like human cognitive process.Convolutional neural network has a great advantage in image target recognition.The excellent performance of convolutional network with different structures makes it the core of computer vision research.It is worth further exploring whether it is the breadth research or depth research in this field.The core research object of this paper is air target recognition,and the application background is to recognize the air obstacle target in front of the UAV.When flying at low altitude(10-100m),four rotor UAV will encounter a variety of targets that affect flight safety,such as tree branches,mountains,buildings,other flying objects,etc.The characteristics of these air targets are different from those of the ground targets.There is no special dataset for target recognition in the high altitude perspective,and the air targets can not be trained with the images taken from the ground.In this paper,aiming at the lack of air target dataset,the air target image dataset(ATI)is designed.The dataset uses UAV aerial photography to collect materials,including four kinds of air obstacle targets and two kinds of other targets that affect flight safety.On this basis,this paper uses the deep convolution neural network for target recognition,and proposes the large target detection network(LTDnet),which uses the feature extraction network composed of five convolution layers plus the max-pooling layer.In the recognition stage,the IOA(Intersection Over All)prediction frame evaluation unit proposed in this paper has higher operation efficiency.On the basis of the above research results,this paper uses the airborne Jetson TX2 platform to test.The test results show that the ltdnet network proposed in this paper can recognize the air obstacle targets in real time after ATI dataset training,and recognize 31 pictures per second,achieving the expected purpose.