Research On Target Detection Method Based On Pruning Strategy For Remote Sensing Image

In recent years,with the development of related technologies,remote sensing images have higher resolution and more details,which puts requirement on the detection capability of target detection model.The output of target detection facilitates tasks such as urban planning,environmental protection,crop monitoring,flooding and fire prevention.While the complex background,noise,weather and illumination intensity make the target detection face more challengesTarget detection is the most basic task of image interpretation.In recent years,Convolutional Neural Network(CNN)performs well in computer vision.Convolutional neural network is inspired by convolution operations in image processing at the beginning of design,so CNN performs better when dealing with image problems.One of the characteristics of deep learning technology is that it requires a large amount of training data,and the development of remote sensing technology brings us many high-resolution images,which makes it possible to apply deep learning to target detection for remote sensing image.However,high-resolution remote sensing images have different characteristics compared with ground-shot images:(1)the background is complex because the image contains a wide area with various objects.(2)the target is usually small and the background is large.For an image that has millions of pixels,the target may be has hundreds of pixels.(3)due to the single shooting angle there are small changes in the appearance of targets which usually rotate in large angle.(4)the difference of ground sample distance makes the different scales of targets.At present,many works used deep features in target detection of remote sensing image while they did not fully consider the specific characteristics mentioned above and the defects of CNN in detecting remote sensing image:(1)CNN does not have rotation invariance.The pooling layer can make slight rotation of the target has no effect on the detection result,while large angle rotation can affect the result.However,the target has strong rotation on remote sensing image.(2)From the lower layer to upper layer,the size of feature map will gradually decrease and we will lose accurate localization information.While targets in remote sensing images are really small,locating targets is more dependent on the information in the lower layer feature map.(3)CNN has large parameters and is computationally intensive.Although the number of high-resolution images is increasing with the development of related technologies,the training set of remote sensing images is still relatively small compared to the ground shooting images.Therefore,it is necessary to consider how to match the size of the network and the training set,and how to reduce the amount of calculation.In addition,in the stage of extracting candidate regions,many works in remotes sensing filed still use Edge Boxes,selective search and so on.The disadvantages of these methods are:(1)use hand-craft features,while hand-craft features are less expressive.(2)it takes too much time to extrac candidate regions.To solve the above problems,we propose a Pruning Strategy based Target Detection for Remote Sensing Image(PSTD)method.In order to improve the computational efficiency,in this paper we propose to build a network architecture using autonomous learning.Specifically,we use pruning strategy to prune the network and then use the pruning network method to target detection network for speeding up the computing of the network.At the same time,we process the original training set so that the network can keep the balance between learning the information of the targets and background.This paper is mainly composed of the following parts:(1)Background review of remote sensing image and target detection.Firstly,we expound the development of the target detection model in recent years,and introduce the typical target detection model from two-stage to end-to-end and the advantages and disadvantages of each model.Then we analyze the imaging technology of remote sensing images and introduce the common methods of processing remote sensing images.(2)We propose Sparse CNN algorithm.We first explain the redundancy of convolutional neural networks and the redundancy is caused by the mismatch of the number of network parameters and the size of datasets,so we propose sparse convolution neural network(Sparse CNN)algorithm.(3)We propose PSTD method.Considering the characteristics of relatively small training dataset of remote sensing image,the sparse convolutional neural network algorithm is used to design the training algorithm to train the target detection network,so as to obtain a faster target detection network which is more matched with remote sensing dataset.Then use the multi-layer feature map to solve the problem of small target and different scale of remote sensing image.For the characteristic that the background of the remote sensing image occupies a large part of the image,we use balance sampling and hard negative mining to increase the expressive power of the network.Finally,we use data augmentation to solve the problem of the strong rotation of the target in remote sensing image.(4)To verify the performance of the proposed target detection method for remote sensing image,we used precision-recall curve,average running time,average precision and mean average precision to design the experiment.To evaluate the pruning ability of the Sparse CNN,we design an another experiment.The experimental results show that Sparse CNN can prune the network to a large extent and PSTD has strong detection ability in remote sensing image dataset.