Aircraft Detection In High-resolution Optical Remote Sensing Images Based On Lightweight Method

Convolutional neural networks(CNN)are widely used in scene classification,target detection,and image segmentation.With the development of software and hardware,network performance has greatly improved,while its scale has also increased exponentially.The storage problems of super-large models and the speed bottleneck of model predictions limit the scope of their applications.The lightweight convolutional neural network target detection algorithm has obvious advantages in scenarios with limited computing and storage conditions and high requirements for prediction speed,and has a wide range of application prospects.The lightweight network remote sensing image aircraft detection module,as one of the components of the remote sensing image intelligent analysis system,can be flexibly embedded in diversified reconnaissance and command platforms,and can play an important role in the field of battlefield situation analysis and precision guidance.This module is also very easy to migrate to other high-value target detection tasks from high-resolution remote sensing images.There are two difficulties in high-resolution optical imaging aircraft detection: one is that there are many types and the samples are small,and the other is that the similarity between different types of models is high.In order to solve the above problems,based on the analysis of the deep learning target detection algorithm and the lightweight method of deep convolutional networks,this paper designs a lightweight aircraft target detection algorithm and lightweight aircraft target recognition algorithm.The main work of this paper has the following three points:(1)Based on the analysis of convolution and feature maps,a detailed analysis and research on the lightweight method of convolutional neural networks was carried out,and the model lightweight method was elaborated in detail.Then,the performance of various lightweight networks is compared on the remote sensing image scene classification data set(2)Aiming at the problem that the general target detection model is large and consumes computing and storage resources,an aircraft target detection algorithm based on lightweight YOLOv4 is proposed.Based on the YOLOv4 multi-scale target detection framework,this algorithm uses a variety of lightweight methods to compress model parameters and increase the calculation speed.Use deep separable convolution and lightweight feature extraction modules to build a deep network,greatly compressing model parameters while speeding up model calculations.Use the feature enhancement module(FEM)to connect different convolution branch expansion features in parallel to obtain a larger receptive field while enhancing the expression ability of the feature on the channel.The residual fusion module(RFM)is used to fuse features of different scales,which can effectively fuse different semantic information while reducing the amount of calculation.Lightweight classification and regression decoupling modules allow different branches to focus on different tasks to improve target detection accuracy.The lightweight YOLOv4 aircraft target detection algorithm can adapt to the characteristics that the proportion of aircraft targets in remote sensing images is small and dense,and the size of different types of aircraft changes greatly.Compared with the classic target detection algorithm,the algorithm model in this paper has fewer model parameters,faster calculation speed,and high accuracy at the same time.(3)Aiming at the problem that the sample size of aircraft models is small,the large model is easy to overfit,and the small model is insufficiently generalized,a dual lightweight network aircraft target recognition algorithm based on the mixture of cropping features is proposed.The algorithm uses a combination of mixed sample data amplification(MSDA)and ensemble learning methods to solve the problem of aircraft target recognition for small-sample lightweight models.For the problem of the small number of aircraft samples,data amplification is incorporated into the convolution calculation,and the feature mixture of different samples is cut to increase the diversity of the data.The network needs to distinguish the aircraft category from the cropped part,and the drive model distinguishes the target from the small features.For the problem of low recognition rate of a single lightweight network model,a dual lightweight integrated network with two inputs and two outputs is adopted.The two networks are integrated through the cropping mixing module and the shared convolution module,and they are distinguished by different coding layers and classification layers.It not only reduces the parameters of the model,but also makes different networks distinguishable.The network’s confrontational cooperation helps to accurately classify aircraft targets with small differences in category.Experiments show that the enhancement effect of this algorithm on aircraft targets is better than other mixed sample data amplification methods,and the performance of dual lightweight network integration is better than that of single network.