Framework For Aircraft Target Detection And Fine-grained Classification

Target detection and recognition is the basic research topic in the field of computer vision,and it is also a prerequisite for a large number of advanced tasks.It has always been a hot research problem in this field.As the application scene is usually complex and varied,need to consider the angle of view changes,light changes,blocking,scale,noise and so on various factors,the same type of object in the image performance is often very different.Human vision can be easily obtained from the image of the high-level semantic information,such as whether the image contains a certain object,where the specific location,which is difficult for the computer,so in the past few decades inspired a large number of researchers close Attention and put into research.In this paper,the performance of the target detection algorithm based on local feature and the target detection algorithm based on the depth learning is realized and compared with the sub-task of the aircraft target detection in the large-scale military airport satellite image.Then,for the sub-task of the aircraft type recognition The performance and advantages and disadvantages of the two classification algorithms based on convolution neural network.Based on the local feature detection,the Hough Forest algorithm based on SIFT feature is used.Firstly,the sparse SIFT feature is extracted on the training set,the Hough Forest model is trained with the category information and the offset.Then,on the test set,The feature descriptor performs the probability Hough vote in the Hough space to search for the location where the target center may exist.The latter is to take the current relatively deep understanding of the depth of the idea of the first target detection task as a separate space bounding-box regression problem,through a 24-layer convolution neural network model to achieve the bouding-Box of prediction.From the performance of the two strategies in the same self-mining data set,the traditional target detection and recognition algorithms on large-size images are often difficult to break through the efficiency,and the convolution neural network algorithm is fully utilized Calculate the advantages of hardware,greatly reducing the time-consuming tasks,while the detection accuracy is also a great degree of improvement.Therefore,relative to the traditional detection algorithm,the depth of learning in the field of target detection there is a huge advantage.Based on the Caffe framework,the image classification network based on AlexNet and the image classification network based on ResNet are realized respectively.Then,the Top-1 error rate and the computational efficiency of the two networks are compared.The classification of these two kinds of network performance differences and the possible causes.