Design Of Multi-target Detection Method For Remote Sensing Image Based On Deep Neural Network

Lots of aerial remote sensing images have been generated with the mature of satellite and unmanned aerial photography technology which is meaning for the research on remote sensing images.The multi target detection function of remote sensing images has great significance and application prospect.It can apply to the construction of smart city&monitor of urban road and development of national security.Deep neural network has achieved very high accuracy for target detection in common scenario area by the exploded development of deep learning technology.However,the current network of target detection is not able to achieve rapid and accurate detection of small target,especially in the high resolution remote sensing images,every target object is nearly a small target.Meanwhile in the real-time requirement of multi target detection for remote sensing image,not only need the ability to detect the small target,but also fast detection speed to ensure the instantaneity and fluency of monitoring screen.To resolve the issue which the target object in remote sensing image is too small to detect,and make sure the target detection speed,this article will be researched by SSD algorithm.This article firstly states the status of detection and explain the difference between remote sensing image and ordinary image.In the meantime it introduces the base theory of target detection,classic target network and apply to multi-object detect on remote sensing image,then get baseline mAP,and visualize convolution feature map to analysis the insufficient of network structure for multi-object section in remote sensing image.The following innovative works have been done based on those points:First of all,this article puts forward the idea of secondary cutting to process remote image base on its characteristic of wide shooting range,high resolution and large scale,this action will make sure the object in image keep mostly information even over-scaling and to avoid the loss of image when its size not match with the cutting size.Secondly,due to the issue which small target object in remotes image is difficult to detect,this article propose the updated network FD-SSD base on SSD.The random clipping step of SDD network data preprocessing layer will been removed by FD-SDD,and it can combine the low-level feature map with high resolution and the high-level feature map with the high semantic information,also dilated convolution is used to increase the receptive field of the third layer feature map,small targets are predicted by low feature maps with high resolution.At the same time,the top level feature graph of 1×1 will be no longer used to generate the target box.The experiment shows that FD-SSD with secondary cutting has better detection effect and mAP is 31.01%higher than SSD300.Then,this article use the target tracking algorithm in order to deepen the engineering application of deep neural network.The experiment shows that smooth skip frame detection not only accelerates the engineering application,but also guarantees the smoothness and stability of the monitoring screen.Finally,the real time monitoring of vehicle which take photo by aerial photography has been achieved by FD-SDD network and smooth skip frame detection which mentioned in this article.