Research On Weak And Small Target And Aircraft Detection Technology Based On Deep Learning

Object detection is one of the five major tasks of computer vision.In recent years,with the rapid development of computer hardware systems,deep learning-based object detection technology has been improved.The accuracy of object detection has far surpassed traditional methods and even exceeded that of the human eye.Weak and small object detection is an important research direction in the field of object detection,which is widely used in several research fields.In civil applications,it can be used for video surveillance and anomaly detection.In the military,it can be used for precision strike weapon systems,aerospace technology,space debris detection,and unmanned reconnaissance.In particular,it is important for the estimation of the battlefield situation and military decision-making.An important bottleneck in weak and small object detection technology is to accurately distinguish weak and small objects from interference in complex backgrounds.The core difficulties are mainly manifested as follows: First,the object is small,and it has insignificant shape and texture information.The deep neural network can extract only a limited number of weak and small object features,so it cannot exert its powerful feature extraction and learning ability.Second,the contrast between the object and the background in the image is low,and the weak and small object can be overwhelmed by the complex background.Third,there are few labeled samples,and the background in the image is complex.The objects often blend in with the background and are even imperceptible to the human eye.Therefore,it is difficult to obtain a large-scale dataset of weak and small objects,which makes it impossible to train the deep neural network sufficiently and thus cannot achieve satisfactory detection results.Aircraft detection based on remote sensing images is an essential application of weak and small object detection.For existing aircraft detection algorithms,most deep learning-based aircraft detection faces two major challenges: Firstly,the performance of detection methods can be greatly limited with relatively small amounts of labeled samples.Secondly,most existing models are resource-hungry,they improve detection performance by increasing the width,depth,and resolution of the neural network.Reducing the computational complexity of the model can lead to a decrease in detection accuracy.However,the hardware systems of many application scenarios do not support large network models.In addressing the above-mentioned problems,this dissertation conducts related research around weak and small object detection in complex backgrounds and aircraft detection based on remote sensing images.The aim is to further improve the robustness of object detection algorithms and reduce the reliance on training samples.The main work of this dissertation includes the following aspects.1.Temporal Model with an encoding-decoding structure is designed to detect weak and small objects in complex backgrounds.This model requires only a small number of labeled samples.In the training process,Temporal Model just learns the artificial weak and small objects and background of the image sequence.The number of artificial weak and small objects is much more than the actual weak and small objects in the WSO dataset.Temporal Model can effectively learn temporal consistency information in the image sequence(the pixels identified as background by the Temporal Model are temporal consistency),and thus weak and small objects can be detected.2.For the problem that it is difficult to obtain large weak and small object detection datasets,a deep feature-based weak and small object detection named PBM is proposed in this dissertation.It is an unsupervised learning algorithm that does not require any labeled samples for training.PBM consists of two parts: OFS and DFP module.The OFS module uses a neural network to extract the deep features of image sequences,and four unique loss functions are used during the training process to obtain the optimal features with better discriminative and robustness.The DFP module takes advantage of the distinct lowdimensional subspace clustering of deep features at different pixel locations to compress them to obtain a background model.Thus,weak and small object detection results are obtained.PBM exploits the powerful feature representation ability of deep neural networks and leverages the pixel-level modeling and unsupervised learning of traditional methods.Consequently,PBM is able to accurately model background features while overcoming the curse of dimensionality.The experimental results show that Temporal Model and PBM have high detection accuracy,with F1＿score of 29.31% and 61.77% on the WSO dataset.The detection accuracy of the two proposed methods is higher than most of the other unsupervised learning algorithms,among which PBM achieves state-of-the-art detection results.Therefore,the research results of this dissertation have high application value and can be used in engineering applications such as civil and military.3.This dissertation combines the advantages of convolutional neural networks and traditional circle frequency filter algorithms.Then,a lightweight end-to-end aircraft detection algorithm called CGC-NET is proposed.This method can achieve high detection accuracy with small samples.Experimental results show that F1＿score can reach 91.06%,while the model size is only 0.88 M,which is very beneficial for practical engineering applications.The CGC-NET obtains better results than some advanced methods in the experiments.