Remote Sensing Target Detection Based On Fusion Multi-scale Feature Networ

Nowadays we cannot live without the aid of computer vision,for example,we need face recognition when we use card punching machines at work and at the end of the day,and we need machines to recognize the number plates of vehicles when we identify them.Therefore,target detection in computer vision is closely related to human daily life,and target detection is also the cornerstone of some important research in computer vision.Object detection is the identification of the position and class of the desired target in a given image.Remote sensing target detection has gradually developed in recent years and is slowly emerging from the military field into people’s life horizons,such as planning for cities,investigating planning resources,and monitoring urban environments.Remote sensing images are characterized by high resolution,large image scale,dense and complex types of objects and large target scale gaps within the image compared to general natural images.These characteristics lead to poor detection results and detection indexes when general-purpose target detection algorithms are applied to remote sensing image target detection.Therefore,some more applicable target detection network frameworks are needed to be innovatively designed specifically for the target detection of remote sensing images.With the high interest of researchers,and the development of deep learning and convolutional neural networks,several excellent papers on target detection have been published in recent years,proposing many reliable and well-performing frameworks.However,existing networks ignore the fact that when there are multi-scale features in the same image when extracting features,not only will detailed information be lost after convolution to ignore small targets,but some targets will also be judged as negative samples when detected by other features layers.Because the remote sensing image contains a lot of information and the background is complex,it will have a serious impact on the detection of the targets in the image,and the final detection accuracy obtained will easily fail to reach the ideal value.Therefore,this paper investigates how to improve and fuse multi-scale feature maps in a better and more efficient way,and constructs two remote sensing target detection network structures based on fused multiscale feature maps.This paper presents two network structures based on fused multi-scale feature maps,one designed for small targets in remotely sensed images and the other a remote sensing target detection algorithm that unfolds the study for all classes.Both algorithms are two-stage detectors,where in the first stage the network structure is improved to extract features and enrich the detailed information of the multi-layer feature map.Then the second stage gets pre-selected frames through the RPN network and then the final result is obtained through classification regression.The two methods make different improvements.One of them adds a residual branch to the pyramid network to reduce the negative impact of gradient disappearance and to enrich the detailed information of the feature map,and adaptively pools the feature maps at the top of the pyramid and recalculates the weights,and fuses them to obtain a new feature pyramid.Another algorithm convolves the multi-scale feature maps after feature extraction and fuses features of the same scale size to obtain a new feature pyramid,thus reducing the number of cases where erroneous results are obtained from the same feature map due to large differences in scale.We finally conduct experiments on the DOTA remote sensing dataset and the experimental results show that our proposed method performs well on remote sensing target detection.