Research On Object Detection Technology In Optical Remote Sensing Based On Deep Convolutional Neural Networks

Object detection in optical remote sensing images is of great importance both in civilian and military aspects.In civilian applications,high-precision object detection can assist urban traffic management and help cities to plan and construct;in military,high-precision object detection helps to accurately lock intrusion targets and protect national security.Traditional object detection algorithms in optical remote sensing images—edge detection algorithm,threshold segmentation method,visual saliency-based method,and shallow machine learning-based algorithms have been difficult to meet the requirements of object detection in today's optical remote sensing images.With the deep development of deep learning,especially the deep Convolutional Neural Networks in the field of image processing,the use of deep Convolutional Neural Networks for object detection in optical remote sensing images has become an unstoppable trend.Compared to complex and inefficient workflows for object detection by using shallow machine learning: First,specific feature extraction algorithms must be designed according to specific goals and then the resulting features are sent to the classifier for classification,deep learning,especially deep Convolution Neural Network can directly extract the effective features of the object.In addition,deep Convolutional Neural Networks can extract more essential features of the object while shallow machine learning can only extract primary or intermediate features such as texture features,edge features,and so on.At present,the object detection algorithms based on deep Convolutional Neural Networks can be roughly divided into two types: two-stage detectors represented by Faster RCNN,and single-stage detectors represented by YOLO and SSD.Compared to the two-stage object detectors,single-stage object detectors are more advantageous in terms of speed,but it does not perform well in accuracy,especially when detecting small-scale targets.Especially in the YOLO series,the detection accuracy on small-scale targets is particularly bad.While the two-stage detectors are slower than the one-stage detectors,their precision on small objects is much better than that of one-stage detectors.In comparison,the two-stage detectors are more suitable for object detection in remote sensing images.Therefore,we conduct research based on the published internationally recognized optical remote sensing dataset.Based on the Faster RCNN detector,a variety of mainstream deep Convolutional Neural Networks are fine-tuned on two representative datasets for object detection,and the performance of each deep Convolutional Neural Network is comprehensively analyzed.From the detailed experimental analysis,we can understand the advantages and disadvantages of each deep Convolutional Neural Networks,we can understand which structure of deep Convolutional Neural Network is more suitable for object detection in optical remote sensing images,we can understand the focus and shortcomings of direct fine-tuning deep Convolutional Neural Networks to do object detection in optical remote sensing images.At the same time,in order to reduce the storage space required by deep Convolutional Neural Networks,we propose a method of subnet convolution to Initially lightweight the network.Through research,it is found that fine-tune deep Convolutional Neural Networks directly for object detection in optical remote sensing images has many shortcomings.Therefore,we have to propose a series of improvements.First of all,for object detection in optical remote sensing images,the surrounding environment is complex and the interference is very high.Therefore,there are a large number of difficult-to-identify objects when using deep Convolutional Neural Networks for object detection.In the process of training,in order to ensure that the network achieves better results on the entire data set,the characteristics of “difficult samples” are often ignored.In response to this situation,we use the Online hard example mining,which is to use the hard-to-identify objects as the main objects of network training for object detection and recognition,and the Online hard example mining effectively improves the overall capability of the network.Secondly,some objects in the optical remote sensing image are very dense and their scale are very small.The effective features of these targets will be seriously lost during the propagation of the deep Convolutional Neural Networks.For this situation,we propose to improve the resolution to reduce the features' lost.However,directly increasing the resolution of the network structure will lead to the decline of the receptive field and the loss of global information,so we use "hole convolution" to ensure the receptive field.In addition,the features obtained by the deep network contain less positioning information,and the positioning information is very important for object detection in optical remote sensing images.Therefore,we propose a method to combine multi-scale features.The method combines the advanced semantic features of the deep network to facilitate classification and the features of the shallow network such as edges and textures to facilitate positioning.Finally,in order to reduce the storage space required by the deep Convolutional Neural Networks' model and increase the portability of the network,I propose a method to lightweight the network.The experimental results show that the proposed methods are feasible.Compared with the current target detection method based on deep Convolutional Neural Networks,our method has great advantages in accuracy and recall rate.In summary,we have intensively studied how to use deep Convolutional Neural Networks for object detection in optical remote sensing images.Based on the analysis of the advantages and disadvantages of various mainstream network structures in object detection in optical remote sensing images,a series of improvements are proposed to achieve high-precision Our research has deep guiding significance for object detection in optical remote sensing images.