Research On Small Target Algorithm Of UAV Land Object Based On Deep Learning

Vehicle small target detection technology in aerial imagery has always been a hard point in the field of computer vision.The main reasons are as follows:(1)There are few target pixels and too few features;(2)The aerial background environment is complex and is greatly affected by the outside world.At present,many traditional vehicle small target detection algorithms can realize the detection of aerial vehicles in a simple background.However,considering the complexity of background in aerial images and the angular diversity of image targets,traditional target detection algorithms are difficult to target vehicle targets in aerial photography.However,with the continuous improvement of computer software and hardware,the deep learning technology based on massive data drive has attracted more and more scholars' attention.The target detection based on deep learning can automatically extract multi-dimensional features,so that deep learning can be obtained in the field of target detection.Widely used.In this paper,a small target detection method for UAV ground vehicle based on improved FPN and Faster R-CNN is proposed.The Faster R-CNN model combined with FPN structure(FFRCNN model)is improved and the aerial image acquired in the field is utilized.After testing,the final improved algorithm model has a 19.7% improvement in the mAP indicator.The main research contents of this topic are as follows:Firstly,the deep learning and convolutional neural networks are summarized.The SSD model in the One-stage model and the Faster R-CNN model in the Two-stage model are selected for small target detection on the aerial vehicle dataset of this paper.Combining with the detection effect,the article selects the Faster R-CNN model to solve the accuracy problem of small target detection.This paper also expounds the small target detection technology in deep learning,and finally chooses to improve with the FFRCNN model as the prototype.Secondly,the data set is developed from two aspects: data collection and data preprocessing.On the one hand,the selection and calibration of the data set are introduced.On the other hand,a combination of Gaussian filtering and bilateral filtering is used to perform filtering operations on the noisy images in the data set.At the same time,in order to solve the problem of over-exposed images,various image enhancement techniques are used to adjust and compare to obtain the optimal mode.Thirdly,according to the research purpose of the subject and the actual situation of the data set,three improvements to the FFRCNN model are proposed:(1)using multi-feature fusion to enhance the receptive field of the network;(2)redefining the anchor information to enhance the adaptability of the network to the target;(3)improving the loss function increases attention to difficult casesFinally,the improved algorithm and the original algorithm are compared to the detection effect.The experimental results show that the improved algorithm takes 0.24 seconds to detect each image.Although the average detection time per image is increased by 0.06 s,the mAP is 93.7%,increases 19.7%.The improved algorithm greatly reduces the network's missed detection rate and false detection rate for small targets,effectively improves the network recognition rate,and the algorithm is robust.