Research On Object Detection Based On Deep Convolutional Neural Network

Object detection based on deep convolutional neural network is a hotspot in recent years.It has a wide range of applications in the security,transportation,medicine,military and other fields.The detection of remote sensing images has great research value in both military and civilian fields.In this paper,the object detection method is deeply studied.On the background of remote sensing image recognition,this paper proposes a method of location and detection based on deep convolutional neural network.The method is regard to a two-step detector structure,but it uses the residual network to extract features,which simplifies the difficulty of learning and avoids the saturation of classification accuracy caused by the increase of network depth.The generation method of region proposal is also improved.We increase the number of anchors in the PRN network,making the model more suitable for the detection of remote sensing aircraft images and use the soft-NMS method instead of the original NMS method to improve the detection performance in the presence of objects that are occluded.In this paper,the remote sensing aircraft objects in the RSOD dataset are used for experiments.The dataset is modified and inverted using the method of flipping and rotating.In the training phase,a training method for sharing the convolutional layer between the RPN part and the Fast R-CNN part in the Faster R-CNN model is designed.Experiments were carried out on the trained network model to quantitatively evaluate the detection performance on different remote sensing data sets.In this paper,the control experiments under different anchor numbers and different object location networks are also designed.The final detection accuracy reached 89.72%,indicating that the algorithm can effectively achieve the object detection task for remote sensing aircraft images.Motion object detection is a problem of great research value.This paper also proposes a moving object detection method.The improved frame difference method uses the interval frame extraction image to recombine the image frame sequence to detect the video target.And then,the remote sensing image detection model proposed in this paper is used to detect the image in single frame video to improve the detection accuracy.Finally,the detection accuracy on the test video reached 85.49%,and the average detection time per frame was 2.563 seconds,indicating that it is feasible to apply the Faster R-CNN model to satellite video data processing,however,there is still a lot of room for development in terms of real-time performance.